From Expert Finding to Entity Search on the Web

From Expert Finding to Entity
Search on the Web
Full-day Tutorial at ECIR 2012
1st April 2012
Gianluca Demartini, Peter Mika,
Thanh Tran, Arjen P. de Vries
http://diuf.unifr.ch/main/xi/EntitySearchTutorial

Presenters
• Dr. Gianluca Demartini
– eXascale Infolab, University of Fribourg, Switzerland
– Research Interests:
• Entity Search
• IR Evaluation
• Semantic Web
01 Apr 2012 2
ECIR 2012 Tutorial - From Expert Finding to
Entity Search on the Web

Presenters
• Dr. Peter Mika
– Senior Researcher, Yahoo! Research, Barcelona
– Semantic Search group at Yahoo! Barcelona
– Semantic Search, Web Object Retrieval, Natural
Language Processing
01 Apr 2012 3

Presenters
• Dr. Thanh Tran
– (Institut AIFB, Universität
Karlsruhe, Germany)
– Semantic Search group at AIFB
– Semantic Search, Semantic Data
Management, Linked Data
Query Processing
01 Apr 2012 4

Presenters
• Prof.dr.ir. Arjen P. de Vries
– Interactive Information Access research group,
Centrum Wiskunde & Informatica (CWI); Delft
University of Technology; Spinque
– Research interest: the intersection between
information retrieval and databases
01 Apr 2012 5Van Rijsbergen, 1979

Entity
• An entity is a “proper noun”, “something that
is referred to”
Outcome of “definition” discussion reported in SIGIR Workshop Report
on The First International Workshop on Entity-Oriented Search (EOS),
Krisztian Balog, Arjen P. de Vries, Pavel Serdyukov, Ji-Rong Wen, ACM
SIGIR Forum, Vol. 45, No. 2, Dec. 2011, pp 43-50
01 Apr 2012 6

Entity Search
• All those search tasks that aim at retrieving as
answer to a user query an entity instead of a
document
– People, Countries, Movies, Restaurants, etc.
01 Apr 2012 7

Motivation
• Information is entity-centric
• Search for information is often conducted
around entities (Query log analysis)
– Many queries (50%) search for specific entities
instead of documents [Kumar&Tomkins09]
• Traditional search retrieves a list of blue links
• Novel web experiences may be designed
around entities instead
01 Apr 2012 8

Here, for one search query “Nicole Kidman”,
various entities make up the answer:
bio
photos
movies
trivia
quotes
(...)
01 Apr 2012 9

Entity-centric Applications
• Enterprise applications
• News portals
• Movie portals
• Product reviews
• Search Engines
01 Apr 2012 10

Entities in SERP
01 Apr 2012 11

Entities in SERP
01 Apr 2012 12

Entity Search: The Pipeline
• Entity Representation (DB/SW)
• Entity Extraction (NLP)
• Entity Linking and De-duplication (DB/SW)
• Entity Storage and Indexing (DB/SW)
• Entity Search and Ranking (IR)
• Result presentation (HCI)
• Evaluation (HCI/IR)
01 Apr 2012 13

Outline
• FULL DAY Tutorial (sorry, this is not a joke :)
• Morning
– Data (Peter)
– Data Management (Thanh)
• Afternoon
– Search and Ranking (Gianluca & Thanh)
– Evaluation (Arjen)
01 Apr 2012 14

Morning
• Data
– Structured vs. Unstructured:
– Entity Profiles: data models, entity identifiers,
standards
– Datasets (Desktop, Enterprise, Wikipedia, Web, RDF)
• Data Management
– Entity Extraction
– Entity de-duplication / data fusion
– Entity storage & indexing
01 Apr 2012 15

Afternoon
• Search and Ranking
– Expert Finding models
– Entity Ranking in Wikipedia
– Web Entity Retrieval
– Entity Search over Structured Data
– Relational Entity Search over Structured Data
• Evaluation
– TREC Enterprise
– INEX Entity Ranking
– TREC Entity
– SemSearch, Ad-hoc Object Retrieval
01 Apr 2012 16

Data
• Web data
– Information Extraction
– Semantic Web
• Non-web data
– Enterprise data
– Desktop data
– Email
– ...
01 Apr 2012
18

Data on the Web
• Most web pages on the Web are generated from
structured data
– Data is stored in relational databases (typically)
– Queried through web forms
– Presented as tables or simply as unstructured text
• The structure and semantics (meaning) of the
data is not directly accessible to search engines
• Two solutions
– Information Extraction [see Part 2]
– Relying on publishers to use Semantic Web formats
• Linked Data vs. metadata in HTML
01 Apr 2012
19

Semantic Web
• Sharing structured data across the Web
– Standard graph-based data model
• RDF
– A number of syntaxes (file formats)
• RDF/XML, RDFa
– Powerful, logic-based schema languages
• OWL, RIF
– Query languages and protocols
• HTTP, SPARQL
01 Apr 2012
20

Resource Description Framework
(RDF)
• Each resource (thing, entity) is identified by a URI or
otherwise it’s a blank node
– URIs are globally unique
• Data is broken down into individual facts
– Triples of (subject, predicate, object)
• A set of triples is published together in an RDF document
example:roi
“Roi Blanco”
name
type
foaf:Person
RDF document
01 Apr 2012 21

An RDF graph
peter#123
“Peter Mika”
name
foaf:Person
sameAs
peter#456
worksWith
roi#234
“roi@yahoo-inc.com”
email
type
type
01 Apr 2012 22

OWL, the Web Ontology Language
• The schema language for the Semantic Web
– Classes, properties and restrictions on their usage
– Allows validation and inference
• Schema is also data
– Published just like any other RDF document
– Queries can refer to both schema and data
• e.g. taxonomy expansion: retrieve instances of a class
and instances of all subclasses
01 Apr 2012
23

Publishing RDF and OWL
• Linked Data
– Data published as RDF documents linked to other RDF
documents
• Typically RDF/XML or Turtle
• Keep an eye on JSON-LD
– Community effort to (re-)publish open datasets
• Embedded metadata
– RDFa, microdata, microformats annotations inside webpages
– Recommended for site owners by Yahoo, Google, Facebook
• SPARQL endpoints
– Triple stores (RDF databases) that can be queried through the
web
01 Apr 2012
24

Linked Data
• Interlinked datasets on the Web
– Often data from existing databases or APIs
• The four rules of Linked Data:
– Use URIs to identify things.
– Use HTTP URIs so that these things can be referred to
and accessed by people and crawlers.
– Use standard formats such as RDF to provide useful
information about the thing when its URI is accessed
– Include links to other datasets
• Most importantly: links to entities in other datasets that
describe the same entity
01 Apr 2012
25

Peter’s homepage
Yahoo!
Friend-of-a-Friend ontology
Linked Data
peter#123
“Peter Mika”
name
foaf:Person
sameAs
peter#456
worksWith
roi#234
“roi@yahoo-inc.com”
email
type
type
01 Apr 2012 26

Linked (Open) Data = LOD
• Advantages:
– No change to the publishing of the HTML documents
– Data can be published by third party (e.g. Dbpedia)
• Disadvantages:
– Web servers need to be configured to properly handle URIs that
identify concepts instead of documents
– Not favored by search engines
• Lack of use cases
• Crawling needs to be changed
• Authority is difficult to determine
• Tools
– Triple stores (Virtuoso, Oracle etc.) and front-ends (Pubby)
– RDB-to-RDF mappers (e.g. D2RQ, Triplify)
– Validators (Vapour)
– Linked Data browsers (many)
01 Apr 2012
27

Linked Data community
• Community effort to (re)publish open datasets
as Linked Data
– In particular, scientific and government data
– see linkeddata.org and ckan.org for developer
information and datasets

Linked Data in practice
• Fetching data dumps
– See catalogs such as thedatahub.org, linkeddata.org
• Crawling Linked Data
– Similar to HTML crawling, but the the crawler needs to
parse RDF/XML (and others) to extract URIs to be crawled
– Semantic Sitemap/VOID descriptions
– Existing crawls
• Billion Triples Challenge (2009-2011) datasets
• LOD cache
• Querying SPARQL endpoints
– See catalogs such as thedatahub.org, linkeddata.org
– Semantic Sitemap/VOID descriptions
01 Apr 2012
29

Datasets
• Broad coverage datasets are linking hubs
– Dbpedia
– Freebase
– Starting in 2012: Wikidata
• Domain-specific datasets form clusters
– Biology
– Government
– Library
– Entertainment
– …
01 Apr 2012
30

Dbpedia
Using
the Dbpedia
ontology
Raw data
32

Metadata in HTML
• 1995: HTML meta tags
• 1998: RDF/XML
• 2003: Web 2.0
– Tagging
– Microformats
– Metadata in Wikipedia
– Machine tags in Flickr
• 2005: eRDF
• 2008: RDFa 1.0
• 2011: RDFa 1.1
• 2012: Microdata
01 Apr 2012
33

HTML meta tags
<HTML>
<HEAD profile="http://dublincore.org/documents/dcq-html/">
<META name="DC.author" content="Peter Mika">
<LINK rel="DC.rights copyright"
href="http://www.example.org/rights.html" />
<LINK rel="meta" type="application/rdf+xml" title="FOAF"
href= "http://www.cs.vu.nl/~pmika/foaf.rdf">
</HEAD>
…
</HTML>
01 Apr 2012
34

Microformats (μf)
• Agreements on the way to encode certain kinds metadata in HTML
– Reuse of semantic-bearing HTML elements
– Based on existing standards
– Minimality
• Microformats exist for a limited set of objects
– hCard (persons and organizations)
– hCalendar (events)
– hResume
– hProduct
– hRecipe
• Varying degrees of support and stability
– hCard and rel-tag are widely supported
• Community centered around microformats.org
– Specifications and discussions are hosted there
01 Apr 2012
35

Microformats: limitations
• No shared syntax
– Each microformat has a separate syntax tailored to the
vocabulary
• No formal schemas
– Limited reuse, extensibility of schemas
– Unclear which combinations are allowed
• No datatypes
• No namespaces, unique identifiers (URIs)
– no interlinking
– mapping between instances is required
• Always appears in the HTML <body>
01 Apr 2012
36

Example: the hCard microformat
<cite class="vcard">
<a class="fn url" rel="friend colleague met” href="http://meyerweb.com/">
Eric Meyer</a> </cite> wrote a post (<cite>
<a href="http://meyerweb.com/eric/thoughts/2005/12/16/tax-relief/">
Tax Relief</a></cite>) about an unintentionally humorous letter he received from
the <span class="vcard”> <a class="fn org url" href="http://irs.gov/">
Internal Revenue Service</a>
</span>.
<div class="vcard">
<a class="email fn" href="mailto:jfriday@host.com">Joe Friday</a>
<div class="tel">+1-919-555-7878</div>
<div class="title">Area Administrator, Assistant</div>
</div>
01 Apr 2012
37

RDFa
• W3C standard for embedding RDF data in HTML documents
– A set of new HTML attributes to be used in head or body
– A specification of how to extract the data from these attributes
• RDFa is just a syntax, you have to choose a vocabulary separately
• RDFa 1.0 is a W3C Recommendation since October, 2008
– RDFa Primer
• RDFa 1.1 currently under standardization
– RDFa Core & RDFa Lite Working Draft as of January 31, 2012
– Updated version of the RDFa Primer
• RDFa API for accessing RDFa data in a webpage in the browser from
JavaScript
– Currently Working Draft (April 19, 2011)
01 Apr 2012
38

RDFa 1.1
• Changes
– New vocab attribute to define the default
namespace for the document or subtree
– Syntax changes for ease of use
– RDFa Lite profile
• RDFa 1.1 is backward compatible with RDFa
1.0
– RDFa 1.1 is recommended if you want to use
HTML5
01 Apr 2012
39

Microdata
• Currently under standardization at the W3C
– Working Draft (May 25, 2011)
• Microdata vs. RDFa
– Microdata is simpler to author
– Lacking some extension features such as co-typing
• HTML5 also has a number of “semantic”
elements such as <time>, <video>, <article>,
<section>…
01 Apr 2012
40

Microdata example
<div itemscope itemid=“http://www.yahoo.com/resource/person”>
<p>My name is <span itemprop="name">Neil</span>.</p>
<p>My band is called
<span itemprop="band">Four Parts Water</span>.
I was born on
<time itemprop="birthday" datetime="2009-05-10">May 10th 2009</time>.
<img itemprop="image" src=”me.png" alt=”me”>
</p>
</div
01 Apr 2012
41

Example: Facebook’s Like and the
Open Graph Protocol
• The ‘Like’ button provides publishers with a way to
promote their content on Facebook and build communities
– Shows up in profiles and news feed
– Site owners can later reach users who have liked an object
– Facebook Graph API allows 3rd party developers to access the
data
• Open Graph Protocol is an RDFa-based format that allows
to describe the object that the user ‘Likes’
01 Apr 2012
42

Example: Facebook’s Open Graph
Protocol
• RDF vocabulary to be used in conjunction with RDFa
– Simplify the work of developers by restricting the freedom in RDFa
• Activities, Businesses, Groups, Organizations, People, Places, Products and
Entertainment
• Only HTML <head> accepted
<html xmlns:og="http://opengraphprotocol.org/schema/">
<head>
<title>The Rock (1996)</title>
<meta property="og:title" content="The Rock" />
<meta property="og:type" content="movie" />
<meta property="og:url" content="http://www.imdb.com/title/tt0117500/"
/>
<meta property="og:image" content="http://ia.media-
imdb.com/images/rock.jpg" /> …
</head> ... 43

Fragmentation of web markup
• Multiple schemas
– Yahoo!’s SearchMonkey – June, 2008
– Google announces Rich Snippets – June, 2009
• Faceted search for recipes – Feb, 2011
– Facebook’s Open Graph Protocol – April, 2010
• ‘Verbs’ added to OGP – September, 2010
– Bing tiles – Feb, 2011
• Different syntax
– Microformats, RDFa, microdata
01 Apr 2012
44

Schema.org
• Agreement between Bing, Google, and Yahoo on
what markup webmasters should use
– Help adoption by reducing fragmentation
– Pre-competitive: each party will continue to build
competing products independently
• Schema.org covers areas of interest to all three
parties
– Business listings (local), creative works (video),
recipes, reviews
– Expected to open up also to external contributions for
non-core areas
01 Apr 2012
45

Example: schema.org
01 Apr 2012
46

Embedded metadata in practice
• 5-10% of webpages contain some explicit
metadata
– Statistics computed from commoncrawl.org give
different results
• Schema.org helped resolve fragmentation
– Except Facebook
• RDFa and microdata likely to co-exist
01 Apr 2012
47

Enterprise Data
• Unstructured
– Technical reports, Product Specification, etc.
• Semi-structured
– E-mail, Spreadsheets
• Structured
– Databases, Repositories
01 Apr 2012
49

Enterprise Search
• Challenges
– Deal with data and format diversity
– Index/search diverse datasets
• Vertical vs Centralized systems
– Deal with security and access control
– Specific informational needs
• Expert Finding
• Writing an overview
01 Apr 2012
50

Desktop Data
• Textual
– Unstructured
• Txt documents
– Semi-Structured
• E-mails, PDFs, Word files, etc. contain much metadata
• Multi-media
– Pictures, Videos, Audio
– Metadata
01 Apr 2012
51

Desktop Search
• IR techniques over unstructured data
• Exploit
– the structure and metadata available
– user activity logs (browsing history, file access
patterns, etc.)
• Beagle++
– Hybrid search over inverted index and RDF store
– E-mail context and attachments
– Folder structure
– Browser cache Enrico Minack, Raluca Paiu, Stefania Costache, Gianluca
Demartini, Julien Gaugaz, Ekaterini Ioannou, Paul-Alexandru
Chirita, Wolfgang Nejdl: Leveraging personal metadata for
Desktop search: The Beagle++ system. J. Web Sem. 8(1): 37-
54 (2010)
01 Apr 2012

Tutorial Outline
• Morning
– Data (Peter)
• Afternoon
01 Apr 2012 53

Agenda
• Knowledge/Entity Extraction
• Entity Linking
• Entity De-duplication
• Entity Storage & Indexing
… very high-level overview of problems and
solutions!
… see tutorials on the specific problems!
01 Apr 2012
55

Knowledge/Entity Extraction
Source: Tadej Steiner from Jozef
Stefan Institute, Ljubljana, Slovenia

Problem definition
• Knowledge extraction:
– Extracting information from data and
– Adding it to a knowledge base
01 Apr 2012
57

Problem definition
• Information extraction + knowledge
acquisition
(textual) data
extracted
infomation
knowledge
base
Information
extraction
Knowledge
acquisition
01 Apr 2012
58

Information extraction
• From the advent of the WWW, there are huge
quantities of unstructured textual data,
where manual information extraction would
be infeasible
• How to extract information from text
automatically with human-comparable quality
01 Apr 2012
59

Information extraction: early solutions
• Match manually defined patterns against text
• Example:
– Patterns like “Pay ? from ? in favor of ?”
– ATRANS (1986) inter-banking message exchange
01 Apr 2012
60

Knowledge acquisition
• How to transform a world (or domain) model
from existing forms into a computer-friendly
form
– Conceptual knowledge (classes, rules, T-Box) VS.
– Instance information (instance data, resource
descriptions, data records, A-Box)
• Use cases for knowledge bases:
– Answering complex entity search queries /
questions in general:
• “which scientists are also politicians?”
01 Apr 2012
61

Knowledge acquisition
• Constructing a knowledge base is expensive
– The Cyc KB was mostly manually constructed over
the last 20 years
• Coupling information extraction and
knowledge acquistion lets us construct a
knowledge base with no or little human effort
01 Apr 2012
62

Challenges
• Human effort:
– Defining (domain-specific and domain-
independent) extraction patterns
– Especially, in case of bootstrapping approaches:
• Specifying relations
• Construction of training examples
– Maintaining knowledge base consistency
01 Apr 2012
63

Related research areas
• Natural language processing
• Information extraction
• Machine learning
• Knowledge management
 Knowledge extraction tools can be compared
by these perspectives
01 Apr 2012
64

General knowledge extraction tools
• WebKB
• TextRunner
• Cyc
• SOFIE with the corresponding YAGO
knowledge base
• Read The Web
• EntityCube
01 Apr 2012
65

Natural language processing
• Employed by most modern approaches
• Part-of-speech tagging
• Noun phrase chunking, used for entity
extraction
• Abstraction of text
– From: “Slovenia borders Italy”
– To:“noun – verb – noun”
01 Apr 2012
66

Information extraction: entities
• Entity extraction / Named Entity Recognition
– “Slovenia borders Italy”
• Entity resolution
– “Apple released a new Mac”.
– From “Apple”, “Mac”
– To Apple_Inc., Macintosh_(computer)
• Entity classification
– Into a set of predefined categories of interest
– Person, location, organization, date/time, e-mail
address, phone number, etc.
– E.g. <“Slovenia”, type, Country>
01 Apr 2012
67

Some NER tools
• Java
– Stanford Named Entity Recognizer
• http://nlp.stanford.edu/software/CRF-NER.shtml
– GATE
• http://gate.ac.uk/
– LingPipe
• http://alias-i.com/lingpipe/
• C
– SuperSense Tagger
• http://sourceforge.net/projects/supersensetag/
• Python
– NLTK
• http://www.nltk.org
01 Apr 2012
68

NER – list lookup
• Entities stored in lists (gazetteers)
– E.g., Countries and cities
• Plus: Simple, fast, cross-language
• Minus: list update, name variants (UPF,
Universitat Pompeu Fabra), ambiguity
01 Apr 2012
69

List lookup – ambiguities
• Term level
– E.g. capitalized words: [All American Bank] vs. All
[State Police]
• Structure level
– “[Dolce and Gabbana]” vs “[Microsoft] and [Yahoo!]”
• Type level
– John Smith (organization vs. person)
– May (person vs. date vs. verb)
– Washington (person vs. location)
– 2015 (date vs. time)
 Gazetteers not end solution but sources of
background knowledge
01 Apr 2012
70

NER methods
• Rule Based
– Regular expressions, e.g. capitalized word + {street, boulevard, avenue} indicates
location
– Engineered vs. learned rules
• NER can be formulated as classification tasks
– NE extraction: assign word mentions to tags (B beginning of an entity, I continues
the entity, O word outside the entity)
– NE classification: assign entity mentions to categories (Person, Organization, etc.)
– Use ML methods for classification: Decision trees, SVM, AdaBoost
– Standard classification assumes cases are disconnected (i.i.d)
• Probabilistic sequence models: HMM, CRF
– Each token in a sequence is assigned a label
– Labels of tokens are dependent on the labels of other tokens in the sequence
particularly their neighbors (not i.i.d).
01 Apr 2012
71

Naïve Bayes Classification
• Determine category of xk by computing for each yi
• Priors P(Y=yi) and conditionals P(X=xk | Y=yi)
estimated from data (via MLE),
– E.g. If ni of the examples in D are in yi then P(Y=yi) = ni / |D|
• When categories are complete and disjoint, P(X=xk):
)(
)|()(
)|(
k
iki
ki
xXP
yYxXPyYP
xXyYP



 




m
i k
iki
m
i
ki
xXP
yYxXPyYP
xXyYP
11
1
)(
)|()(
)|(


m
i
ikik yYxXPyYPxXP
1
)|()()(
01 Apr 2012
72

Classification via Logistic Regression
• Instead of generative models, a descriminative model can be
used to specifically focus on the conditional distribution P(Y | X)
• Assumes a parametric form for directly estimating P(Y | X)
• Basically a linear model
01 Apr 2012
73


 n
i ii Xww
XYP
10 )exp(1
1
)|1(






 n
i ii
n
i ii
Xww
Xww
10
10
)exp(1
)exp(
)|1(
)|0(
1iff0labelAssign
XYP
XYP
Y


 

n
i ii Xww 10 )exp(1


n
i ii Xww 100


n
i ii Xww 10lyequivalentor
)|1(1)|0( XYPXYP 

Classification
Y
X1 X2
… Xn
Y
X1 X2
… Xn
Naïve
Bayes
Logistic
Regression
Conditional
Generative
Discriminative
01 Apr 2012
74

Sequence Labeling
Y2
X1 X2
… XT
HMM
Linear-chain CRF
Conditional
Generative
Discriminative
Y1 YT
..
Y2
X1 X2 … XT
Y1 YT
..
01 Apr 2012
75
Sunita Sarawagi and William W.
Cohen. Semi-Markov Conditional
Random Fields for Information
Extraction. In NIPS, 2005.

NER features
• Gazetteers (background knowledge)
– location names, first names, surnames, company names
• Word
– Orthographic
• initial-caps, all-caps, all-digits, contains-hyphen, contains-dots,
roman-number, punctuation-mark, URL, acronym
– Word type
• Capitalized, quote, lowercased, capitalized
– Part-of-speech tag
• NP, noun, nominal, VP, verb, adjective
• Context
– Text window: words, tags, predictions
– Trigger words
• Mr, Miss, Dr, PhD for person and city, street for location
01 Apr 2012
76

Exploiting Query Logs / Click-Through Data
• Weakly-supervised entity Extraction from
queries / click-through data
– A small set of seed instances for each entity type
• Learn
– Patterns captured by LDA-based topic model:
Probabilities of query contexts and click websites of
named entities for each class
– Template: common query prefix and postfix, e.g.
“how did country gain independence”
• Apply patterns / templates to click-through data /
query logs to mine new named entities
Marius Pasca: Weakly-supervised discovery of named entities using web search queries. CIKM 2007:683-690
Gu Xu, Shuang-Hong Yang, Hang Li: Named entity mining from click-through data using weakly supervised
latent dirichlet allocation. KDD 2009:1365-1374
01 Apr 2012
77

Information extraction: relations
• Relation extraction
– <“Slovenia”, “borders”, “Italy”>
• Relation resolution
– <“Slovenia”, borders, “Italy”>
– <“Slovenia”, next_to, “Italy”>
• We distinguish between open and
bootstraped approaches
01 Apr 2012
78

Relation Extraction
• Extracting relations
– Typical paraphrase problem: identify all the ways a relation may be
expressed
• Formulated as classification task, e.g. uses SVM
– Training data: parse tree, with nodes associate with a type as well as a
role (e.g. role=member, role=affiliation to capture a person-affiliation
relation)
– Tree-based kernel: two trees are similar if roots have same type and role,
and each has a subsequence of children (not necessarily consecutive)
with the same types and roles
– Examples are converted into such parse trees with role labels, and used to
train the system
– SVM can then classify new examples of possible relations
• Formulate as sequence labeling (semantic role labeling)
• Joint inference: considers different types of features (syntactic,
semantic) and problems (extraction, resolution)
01 Apr 2012
79

Bootstrapped information extraction
• Provide examples for relationships which we
want to extract
• Compromise: lower coverage, higher quality
• Example: Sofie, ReadTheWeb
01 Apr 2012
80

Open information extraction
• We do not want to put constraints on the
types of relationships we want to extract
• Very interesting for open-domain WWW
datasets
• Example: TextRunner
• Compromise: higher coverage, lower quality
• Hybrid approaches:
– EntityCube combines both bootstrapped and open
extraction
01 Apr 2012
81

Knowledge management
• Organization
• Consistency management
• Strictness
01 Apr 2012
82

Knowledge organization
• Lexicon: A set of entities and statements
• Ontology: A complex graph of formal concepts
– Not only concrete entities, but also abstract
classes
– Sofie/Yago, WebKB, ReadTheWeb, TextRunner
• Full world model: A Context-sensitive complex
graph
• Cyc
01 Apr 2012
83

Knowledge consistency
• Consistency management
– Not all extracted information is accurate
– Inaccurate information leads to inconsistencies in
the knowledge base
– Example:
• Having pattern “?x is mayor of ?y” and knowledge that
<x,mayorOf,y> requires <x,type,Person> and
<y,type,City>, we can filter out inconsistent
information
01 Apr 2012
84

Knowledge consistency
• Examples:
– SOFIE:
• Select the subset of statements which have the
maximum satisfiability with regard to constraints
– ReadTheWeb:
• Learns new constraints via semi-supervised boostrap
learning
• Accuracy grows with ontology complexity
01 Apr 2012
85

• Bootstrapping
– Using existing manually prepared knowledge to
generate new knowledge
– While the knowledge base grows, the rules for
extraction also change
01 Apr 2012
86

• Strictness:
– When do we consider entity and relationship
resolution important?
• Depends on use case:
– Reasoning and data integration requires well-
formed and unambigouous entities and relations
– Information retrieval can cope with not-well
formed relations and entities
01 Apr 2012
87

Machine learning
• Used in NLP, IE as well as knowledge
acquisition
• Various approaches
– Self-supervised
– Semi-supervised
– Supervised
01 Apr 2012
88

Machine learning
• Natural language processing
– Part-of-speech learning
• Information extraction
– Pattern learning
• ReadTheWeb, TextRunner, WebKB
• Knowledge acquisition
– Rule learning (WebKB)
– Constraint learning (ReadTheWeb)
01 Apr 2012
89

Summary
• Cyc
– Full world model knowledge base
• WebKB
– First attempt of automatically constructing a
knowledge base
• TextRunner
– Open information extraction
01 Apr 2012
90

Summary
• EntityCube
– Hybrid bootstrapped and open IE
• SOFIE/YAGO
– Tight integration of natural language processing,
disambiguation and acquisition
• Read The Web
– Constraint learning
01 Apr 2012
91

Entity Linking
Source: Tadej Steiner from Jozef
Stefan Institute, Ljubljana, Slovenia

Basic situation
01 Apr 2012
93

Pipeline
1. Identify named entity mentions in source
text using a named entity recognizer
2. Given the mentions, gather candidate KB
entities that have that mention as a label
3. Rank the KB entities
4. Select the best KB entity for each mention
01 Apr 2012
94

Pipeline
1. Identify named entity mentions in source
text using a named entity recognizer
2. Given the mentions, gather candidate KB
entities that have that mention as a label
3. Rank the KB entities
4. Select the best KB entity for each mention
01 Apr 2012
95

Linking approaches - pair-wise linking
• Pair-wise linking: for each in-text entity,
choose the candidate entity which is the best
w.r.t. description similarity and textual
features
• Is each disambiguation choice independent?
– Pair-wise vs. collective disambiguation
01 Apr 2012
96

Important ranking features
• Mention popularity – P(entity|mention)
– P(dbpedia:Kashmir_(song)|”Kashmir”) = 0.54
– P(dbpedia:Kashmir_(region)|”Kashmir”) = 0.91
– Distribution of links and anchors in Wikipedia
Context similarity - sim(ctx(mention), ctx(entity))
Context of a mention is the surrounding sentences
Context of an entity is the description of the entity (Wiki article)
Coherence / Collective
Entities that appear together tend to be related to one another
Usually solved by a greedy graph pruning algorithm
01 Apr 2012
97

Collective linking
• For each in-text entity, choose the candidate entity
which is the most similar to the in-text entity and
related to other entities that are already chosen.
01 Apr 2012 98
Tadej Stajner, Dunja Mladenic: Entity Resolution in Texts Using Statistical Learning and
Ontologies. ASWC 2009:91-104
Xianpei Han, Le Sun, Jun Zhao: Collective entity linking in web text: a graph-based
method. SIGIR 2011:765-774

Relatedness
• Intuition: entities that co-occur in the same
context tend to be more related
• How can we express relatedness of two
entities in a numerical way?
– Statistical co-occurrence
– Similarity of entities’ descriptions
– Relationships in the ontology
01 Apr 2012
99

Semantic relatedness
• If entities have an explicit assertion connecting
them (or have common neighbours), they tend
to be related
Elvis
Memphis
Elvis
Presley
Memphis,
Egypt
Memphis,
TN
origin
Person
Location
type
type
type
St. Elvis
type
01 Apr 2012 100

Co-occurrence as relatedness
• If distinct entities occur together more often
than by chance, they tend to be related
Document
FC
Barcelona
Bayern
FC
Barcelona
Bavaria
Bayern
München
Mutual information
Mutual information
x
y
x
y
x
y
z
01 Apr 2012
101

Content similarity as relatedness
• If distinct entities have higher similarity of
their descriptions, they tend to be related
Document
a
b
x
y
z
similarity = 0.35
similarity = 0.25
similarity = 0.35
similarity = 0,7
similarity = 0,1
similarity = 0,2
01 Apr 2012 102

Architecture
Input text
Preprocessing
(entity extraction and
consolidation)
.. with in-
text
entities
Background
knowledge
(ontology)Match
retrieval
Entity description
vectors
Assertion type
informativeness
Entity
co-occurences
.. with
resolved
entities
Relatedness
Entity
linking 103

Crowdsourcing for Entity Linking
Micro
Matching
Tasks
HTML
Pages
HTML+ RDFa
Pages
LOD Open Data Cloud
Crowdsourcing
Platform
Z enCrowd
Entity
Extractors
LOD Index Get Entity
Input Output
Probabilistic
Network
Decision Engine
Micro-
TaskManager
Workers Decisions
Algorithmic
Matchers
Gianluca Demartini, Djellel Eddine Difallah, and Philippe Cudré-Mauroux.
ZenCrowd: Leveraging Probabilistic Reasoning and Crowdsourcing Techniques for Large-Scale
Entity Linking. In: 21st International Conference on World Wide Web (WWW 2012), Lyon,
France, April 2012.
104

• Matching micro-task
– Unclear (i.e., low confidence) matches are
crowdsourced
– Top algorithmic results are presented to the
workers
– Answers from the crowd are input to a
probabilistic network
01 Apr 2012
105

• Probabilistic Graph
– Worker prior probability (from previous tasks)
– Link prior probability (from algo matchers)
– Link factors connect worker clicks and links
– SameAs constraints
– Dataset unicity contstraints
w1
w2
l1
l2
pw1( ) pw2( )
lf1( ) lf2( )
pl1( ) pl2( )
l3
lf3( )
pl3( )
c11
c22
c12
c21
c13
c23
u2-3( )sa1-2( )
01 Apr 2012

Entity De-duplication
“Entity Consolidation”
“Entity Resolution”
“Record Linkage”
“Instance Matching”
Sources: Yongtao Ma from Karlsruhe Institute of Technology, Samur Araujo from
The Delft Bioinformatics Lab and Aidan Hogan from Digital Enterprise Research Institute

Structure
• Motivation
• Problem and task overview
• Consider only explicit owl:sameAs
• Consider some lightweight reasoning
• Inductive / instance matching methods
– Effectiveness
– Efficiency
01 Apr 2012
108

Motivation 340，000
Results
01 Apr 2012
109

Motivation
• 2% of customer records obsolete in 1 month, due to deaths, name
changes
• $611B/year loss in US due to poor customer data
• An average company has 49 different databases and spends 35% of its
IT dollars on integration efforts
01 Apr 2012
110

Motivation
01 Apr 2012
111

Hetereogenity in naming…
Tim Berners-Lee: URIs
…
timbl:i
dblp:100007
identica:45563
adv:timblfb:en.tim_berners-lee
db:Tim-Berners_Lee
= owl:sameAs
01 Apr 2012
112

11
3
De-duplication for Web data
01 Apr 2012
113

Problem and Task Overview

Data integration – big picture
• Ontology matching
– Widely studied in Semantic Web research, see e.g. list of publications
at ontologymatching.org
• Entity de-duplication
– Logic-based approaches in the Semantic Web
– Studied as record linkage in the database literature
– Machine learning based approaches, focusing on attributes
– Graph-based approaches, see e.g. the work of Lisa Getoor are
applicable to RDF data
• Improvements over only attribute based matching
• Blending / data fusion
– Merging objects that represent the same real world entity and
reconciling information from multiple sources
– Information quality / redundance
01 Apr 2012
115

De-duplication
• The problem of determining if two instances refer to the small real-
world entity.
owl:sameas
Source Instances Target Instances
01 Apr 2012
116

1. Find equivalences in the data
• Consider only explicit owl:sameAs (baseline)
• Consider some lightweight reasoning (extended)
• Inductive / instance matching methods
2. Rewrite data according to equivalences (data fusion)
De-duplication – task overview
01 Apr 2012
117

Consider only explicit owl:sameAs

• Use provided owl:sameAs mappings in the data
timbl:i owl:sameas identica:45563 .
dbpedia:Berners-Lee owl:sameas
identica:45563 .
• Store “equivalences” found
timbl:i ->
identica:45563 ->
dbpedia:Berners-Lee ->
timbl:i
identica:45563
dbpedia:Berners-Lee
De-duplication
01 Apr 2012
119

• For each set of equivalent identifiers, choose a
canonical term
timbl:i
identica:45563
dbpedia:Berners-Lee
De-duplication
01 Apr 2012
120

• Afterwards, rewrite identifiers to their canonical
version:
De-duplication
timbl:i rdf:type foaf:Person .
identica:48404 foaf:knows identica:45563 .
dbpedia:Berners-Lee dpo:birthDate “1955-06-08”^^xsd:date .
dbpedia:Berners-Lee rdf:type foaf:Person .
identica:48404 foaf:knows dbpedia:Berners-Lee .
dbpedia:Berners-Lee dpo:birthDate “1955-06-08”^^xsd:date .
timbl:i
identica:45563
dbpedia:Berners-Lee
01 Apr 2012
121

Consider some lightweight reasoning

• Infer owl:sameAs through reasoning (OWL 2
RL/RDF)
1. explicit owl:sameAs (again)
2.owl:InverseFunctionalProperty
3.owl:FunctionalProperty
4.owl:cardinality 1 / owl:maxCardinality 1
foaf:homepage a owl:InverseFunctionalProperty .
timbl:i foaf:homepage w3c:timblhomepage .
adv:timbl foaf:homepage w3c:timblhomepage .
⇒
timbl:i owl:sameas adv:timbl .
…then apply data fusion as before
Extended de-duplication
01 Apr 2012
123

Inductive / Instance Matching
Methods

Agenda
• Problem overview
• Attribute level
– (see term matching)
• Instance level
– Effectiveness: learning
– Efficiency: blocking
• Dataset level
– (see collective entity linking)
01 Apr 2012
125

Problem overview
effectiveness vs. efficiency
Instance Matching
Effectivity
Find correct matches!
Efficiency
Do it fast!
01 Apr 2012
126

Efficiency
O(NxM)
Source Target
Not efficient
01 Apr 2012
127

“Diclofenac” occurrence on DBPEDIA
01 Apr 2012 128

Problem overview – attribute level
<A1, ‘Dave White’, ‘Intel’, ‘Male’> <P1, ‘Database…’, ‘John Black’, ‘Don White’>
<A2, ‘Don White’, ‘CMU’, ‘Male’> <P2, ‘Multimedia…’, ‘Sue Grey’, ‘D. White’>
<A3, ‘Susan Grey’, ‘MIT’, ‘Female’> <P3, ‘Title3…’, ‘Dave White’>
<A4, ‘John Black’, ‘MIT’, ‘Male’> <P4, ‘Title5…, ‘Don White’, ‘Joe Brown’>
<A5, ‘Joe Brown’, unknown, ‘Male’><P5, ‘Title6…’, ‘Joe Brown’, ‘Liz Pink’>
<A6, ‘Liz Pink’, unknown, ‘Female’> <P6, ‘Title7…’, ‘Liz Pink’, ‘D. White’>
Attribute level
‘Don White’ , ‘D. White’
‘Don White’, ‘Dave, White’
• What (values?)
• How
• Similarity metrics
• Similarity threshold
• Matching techniques
01 Apr 2012
129

Problem overview – instance level
Instance level
<A1, ‘Dave White’, ‘Intel’, ‘Male’>
<A2, ‘Don White’, ‘CMU’, ‘Male’>
• What (attributes?)
• How
01 Apr 2012
130

• How
Problem overview – dataset level
Dataset level
• What (instances?)
01 Apr 2012 131

Agenda
• Attribute Level
• Instance Level
• Dataset Level
01 Apr 2012
132

Character-based
• [see term matching in Part 3 on search & ranking]
• Edit Distance [G98]
– Character Operations: insert, delete, replace
– Given two strings, s and t, edit(s,t):
• Minimum cost of operations transforming s to t
• Exp.: edit(Eorror, Eror)=1, edit(great,grate)=2
– Aiming at: common typing errors
– Problem: works not well with other type of errors
• Exp.: D. White vs Dave White
• Jaro Rule
01 Apr 2012
133

Token-based
• Q-gram
– The q-grams are short character substrings of
length q of the string
– Example: 3-gram(White)={ ‘Whi’, ‘hit’, ‘ite’ }
– set similarity then can be applied to the grams
01 Apr 2012
134

Agenda
• Attribute Level
• Instance Level
• Dataset Level
01 Apr 2012
135

Questions
• Given instance attributes {Name, Institute,
Gender, Publish}
– Which ones are more important?
– Which similarity measures should be adopted?
– What is the threshold of similarity that should be
adopted?
01 Apr 2012
136

Bayes Decision Rule
• Notation
– A,B are two tables, of n comparable fields
– tuple pairs
– classes: M (match) and U (non-match)
– random vector , xi shows the level of
agreement of the ith field for
• Decision rule:
01 Apr 2012 137
called likelihood ratio

Bayes Decision Rule
• Given training data, assume p(xi|M) and
p(xj|M) are independent for i≠j[5]
• Extension:
– Using an expectation maximization (EM) algorithm
to estimate likelihood
– Relax independent assumption
– Decision with reject class
01 Apr 2012
138

Agenda
• Attribute Level
• Instance Level
• Dataset Level
01 Apr 2012
139

Blocking strategies
Source Target
• Used to reduce the number of instance comparison
• Non-overlapping partitions
• Canopies and clustering – overlapping partitions
01 Apr 2012
140

Blocking strategies
Blocking
Attribute
dependent
Attribute
agnostic
When the source and
target schema match.
Otherwise
01 Apr 2012
141

Attribute dependent
• Blocking Key Value (BKV)
– Sorted Neighborhood approach
– Q-grams blocking technique
• Blocking keys are highly
discriminating attributes (e.g. last
name, phone number)
• Targeting homogeneous datasets
b
a (blocking key)
c
d
01 Apr 2012
142

Sorted Neighborhood
• Motivation：
– similar records have similar values
– multiple “cheap” passes faster than an “expensive” one
• Goal: sort feature by a key to bring matching
records close to each other
• Methodology:
– Create a key for every record (e.g. first 3 characters of last
name)
– Sort data by the key
– Pair-wise comparison within a small sliding window
– Multiple passes based on distinct key
01 Apr 2012
143

Sorted Neighborhood
• Example:
ID Name SS Birthday ZIP
r1 David Black 123-45 01.05.1985 76137
r2 Dauid Black 123-45 01.06.1985 76137
r3 David White 325-52 23.09.1984 84212
r4 David B. 126-53 30.10.1983 84123
r5 David B. 745-32 07.05.1973 84212
r1
r2
r4
r3
r5
r2
r1
r3
r4
r5
ZIP[1..3]
Name[1..3]
01 Apr 2012
144

Q-gram blocking
• Motivation: similar matches have high overlaps of q-grams
• Goal: relaxes the edit distance constraint to a weaker count
constraint on the number of matching q-grams
• Methodology:
given two strings s and t, and a edit distance constraint k
– Count Filtering: s and t must share LBs,t=max(|s|,|t|)-1-(k-1)*q q-
grams
– Position Filtering: s and t must share at least LBs,t positional q-
grams
– Length Filtering: ||s|-|t||≤k
01 Apr 2012
145

Q-gram blocking
• Example:
3-gram
s=abaxabaaba ##a,#ab,aba,bax,axa,xab,aba,baa,aab,aba,ba$,a$$
t=abaabaaba ##a,#ab,aba,baa,aab,aba,baa,aab,aba,ba$,a$$
ED(s,t)≤k → |Q(s) ∩ Q(t)| ≥ max(|s|,|t|)-1-(k-1)*q
ED(s,t)=1, |Q(s) ∩ Q(t)|=9
01 Apr 2012
146

Attribute dependent
• Learning the attributes (blocking keys)
– Decision tree
– Maximum hyper-rectangles
DrugBank
DBPEDIA
Label
Drugname
Sideeffect
Page
Title
Name
Producer
Composition
01 Apr 2012
147

Attribute dependent
• Learning functions of similarity (e.g., Jaccard, Jaro, Levenshtein, Hamming, Cosine, etc.)
DrugBank
DBPEDIA
Label= TitleDiclofenac Diclofenac Sodium =≈
01 Apr 2012
148

Attribute agnostic
• Designed for heterogeneous
information space. (i.e., loose
schema binding, noise, missing or
inconsistent values, as well as an
unprecedented level of
heterogeneity)
• No knowledge about the schema
software
Corp. (blocking key)
radio
film
01 Apr 2012
149

Attribute agnostic
• “All tokens”
• Reduce comparison space
– Block purging,
– Block scheduling,
– Block enumeration,
– Duplicate propagation,
– Comparisons propagation, and
– Comparisons pruning.
software
Corp. (blocking key)
radio
film
01 Apr 2012
150

Indexing
• Search requires matching and ranking
– Matching selects a subset of the elements to be
scored
• The goal of indexing is to speed up matching
– Retrieval needs to be performed in milliseconds
– Without an index, retrieval would require scanning
through the collection
• The type of index depends on the types of data
and queries to be supported
– DB-style indexing
– IR-style indexing
01 Apr 2012
152

IR-style indexing
• Index data as text
– Create virtual documents from data
– One virtual document per subgraph, resource or
triple
• typically: resource
• Key differences to Text Retrieval
– RDF data is structured
– Minimally, queries on property values are required
01 Apr 2012
153

Horizontal index structure
• Two fields (indices): one for terms, one for properties
• For each term, store the property on the same position
in the property index
– Positions are required even without phrase queries
• Query engine needs to support the alignment operator
• Dictionary is number of unique terms + number of
properties
01 Apr 2012 154

Vertical index structure
• One field (index) per property
• Positions are not required
– But useful for phrase queries
• Query engine needs to support fields
• Dictionary is number of unique terms
• Number of fields could be a problem for merging,
query performance
01 Apr 2012 155

Indexing using MapReduce
• MapReduce is the perfect model for building
inverted indices
– Map creates (term, {doc1}) pairs
– Reduce collects all docs for the same term: (term,
{doc1, doc2…}
– Sub-indices are merged separately
• Term-partitioned indices
01 Apr 2012
156

Outline
• Expert Finding models
• Entity Ranking in Wikipedia
• Web Entity Retrieval
• Entity Search over Structured Data
• Relational Entity Search over Structured Data
01 Apr 2012
158

From Documents to Entities
• Document Search
01 Apr 2012
159

From Documents to Entities
• Entity Search
1. Ent1
2. Ent2
3. Ent3
01 Apr 2012
160

A taxonomy of Entity Search tasks
01 Apr 2012 161

Expert Finding - Motivation
• Scenario
– In large companies competencies and
skills are spread
– Executives need to create a team for a new
project: find staff with the right expertise
– Someone needs to solve a problem
– Example: I need an expert on ontology
engineering
01 Apr 2012
162

Expert Finding - Motivation
• Goal
– Use the digital content available in the
enterprise
– Create a ranking of people who are experts
in the given topic
01 Apr 2012
163

Motivation for System Support
• Busy experts do not have time to maintain
adequate descriptions of their continuously
changing specialized skills
• Expert seekers have poorly articulated
requirements and are not fully enabled to
judge a good expert from a bad one
01 Apr 2012
164

Complicating factors
• Volume of communication/publication is not a
reliable indication of expertise
• Certain topics engender more opinion than facts
• Lack of information about past performance of
experts
• New employees don’t know about informal social
networks
• Access to expertise is often controlled (informally or
formally, by the experts or their management)
• Solutions to complex problems require diverse
ranges of expertise
01 Apr 2012
165

Evidence of Expertise
• Email or bulletin board messages
• Corporate communications
• Shared folders in file system
• Resumes and homepages
• Employee database
• Email flow
• Bibliographic information
• Software library usage
• Search and publication history
• Project time charges
See also bibliography on TREC-ENT wiki:
http://www.ins.cwi.nl/projects/trec-ent/wiki/index.php/Bibliography
Content
Social
networks
Activities
01 Apr 2012 166

Assumptions
• Content
– Experts are mentioned in relevant documents
– Experts author relevant documents
• Social networks
– People that interact are likely to share expertise
– Evidence in records of information exchange (and
co-authorship, co-work) and/or organizational
structure
01 Apr 2012
167

Two Basic Approaches
Who should I ask about the copyright forms?
• Document-based: rank
docs, extract experts
Copyright forms
Lori
Lori
Lori
Ellen
Ian
Lori
Lori
Ellen
Lori
1.
2.
1.
4.
5.
3.
01 Apr 2012
168

Document-based Expert Finding
• Find and score documents about the topic
– Title about topic
– Abstract about topic
• Aggregate scores for each distinct author
01 Apr 2012
169

Two Basic Approaches
Who should I ask about the copyright forms?
• Document-based: rank
docs, extract experts
• Candidate-based: rank
candidate profiles
Copyright forms
Lori
Lori
Lori
Ellen
Ian
Lori
Lori
Ellen
Lori
1.
2.
1.
4.
5.
3.
Lori
Copyright forms
Ellen
Ian
01 Apr 2012
170

Voting model
• Data fusion techniques
• Each ranked document represents a vote for
the expertise of a candidate
• Vote aggregation:
– Number of docs voting for each candidate
– Scores of retrieved documents
– Ranks of retrieved documents
01 Apr 2012
Craig Macdonald, Iadh Ounis: Voting for candidates: adapting
data fusion techniques for an expert search task. CIKM 2006:
387-396

User-Oriented Model
• Additional real-world constraints
• Distance between user and expert
– User previous knowledge on the topic
– Contact time (organizational hierarchy, geo
location, collaboration)
01 Apr 2012
Elena Smirnova, Krisztian Balog: A User-Oriented Model for Expert Finding. ECIR 2011: 580-592
172

Additional Techniques
Research Systems
• Combine the two basic approaches
• Estimate the quality of the evidence
• Use of collection/structural knowledge
– Treat emails different from documents
– Treat email’s subject/sender/receiver different
from body
– Locate homepages
See also TREC proceedings 2005-2007
01 Apr 2012
173

Additional Techniques
Research Systems
• Use social network extracted from co-
authorship or email lists
• Relevance propagation over expertise graph
• Use Web Search evidence
01 Apr 2012
174

Expert Finding - References
– P@noptic Expert [Craswell et al.
Ausweb01]
– Balog’s Model 1 and 2 [Balog et al.
SIGIR06]
– Voting Model [Macdonald and Ounis
CIKM06, ECIR07, ECIR08]
– Expertise evidence [Macdonald et al.
ECIR08]
– Vector Space Model [Demartini et al.
ECIR09]
– Web evidence [Serdyukov et al. TREC08]01 Apr 2012
175

Ranking…
• People
• Actors
• … Car companies
[i.e., insert your fav entity type here]
Entity Ranking!!!
01 Apr 2012
177

Wikipedia
• Encyclopedia
– multilingual, Web-based, free-content, openly-
editable: errors are promptly corrected
• Articles:
– balanced, neutral, and encyclopedic, containing
notable verifiable knowledge
• Categories / sub-categories
• Links, anchor text (Germany -> Albert Einstein)

Entities in Wikipedia
• Art museums
• Countries
• Actors, Singers
• Monarchs
• Artists
• Magicians
• ...
01 Apr 2012
179

Example Entity Ranking Scenarios
• Impressionist art museums in Holland
• Countries with the Euro currency
• German car manufacturers
• Artists related to Pablo Picasso
• Countries involved in WWI
• Actors who played Hamlet
• English monarchs who married French women
Many examples on
http://www.ins.cwi.nl/projects/inex-xer/topics/
01 Apr 2012
180

Entity Ranking
• Topical query Q
• Entity (result) type TX
• A list of entity instances Xs
• An entity is represented by its Wikipedia page
• Systems employ categories, structure, links
01 Apr 2012
181

Tasks
• Entity Ranking (ER)
– Given Q and T, provide Xs
• List Completion (LC)
– Given Q and Xs[1..m]
– Return Xs[m+1..N]
01 Apr 2012
182

Topic 60
Title
olympic classes dinghy sailing
Entities
470 (dinghy) (#816578)
49er (dinghy) (#1006535)
Europe (dinghy) (#855087)
Categories
dinghies (#30308)
Description
The user wants the dinghy classes that are or have
been olympic classes, such as Europe and 470.
Narrative
The expected answers are the olympic dinghy classes,
both historic and current. Examples include Europe and
470.
TX
Q
Xs
01 Apr 2012 183

Formal Model for Entity Ranking
– Indexing
• Entities
• Data Sources
“Alexandre Pato”
ID: ap12dH5a
(born in; 1989)
(playing with; acm15hDJ)

Formal Model for Entity Ranking
• Searching
– Users' Information Need
– Entity Ranking System

Approaches to ES in Wikipedia
• Exploit and refine the category structure
– Wordnet to find entity types (e.g., a professor is a
person)
• Extend the query
– Synonyms and related words (Wordnet synsets)
• Exploit the link structure
– Links in Wikipedia are usually entities
– Search Keywords also in anchor text of outLinks
01 Apr 2012
186

YAGO
– Suchanek et al. 2007
– Highly accurate ontology
(>95%)
– Extracted from Wikipedia +
WordNet
– Provides semantic concepts
describing Wikipedia entities
Married...
With
Children
Sitcoms
WordNet
Synset
Wikipedia
Category
Wikipedia Taxonomy
YAGO subClassOf
Situation
Commedy

Category Based Search
• Query expansion by modifying category
information
– Subcategories
• Extracted from Wikipedia
– “Children” Categories
• Filtered using the YAGO subClassOf relation
– “Sibling” Categories
• Extracted from Wikipedia
• Having with the same YAGO type

Subcategories
Sitcoms
Wikipedia
Subcategories
Latino
Sitcoms
Sitcoms in
Canada
BBC
Television
Sitcoms
Sitcom
Characters
Wikipedia
Category

“Children” Categories
Sitcoms
YAGO subClassOf
Latino
Sitcoms
Sitcoms in
Canada
BBC
Television
Sitcoms
Sitcom
Characters
Situation
Comedy
Fictional
Character
Wikipedia
Subcategories
Wikipedia
Category

“Sibling” Categories
Sitcoms
YAGO subClassOf
Latino
Sitcoms
Sitcoms in
Canada
BBC
Television
Sitcoms
Situation
Comedy
Wikipedia
Category
YAGO
subClassOf
...

Entity Search over Wikipedia
• Search for many different entity types with one
system!
• Main observations
– Link information is important
– Cleaning the category structure of Wikipedia is critical
(YAGO)
– NLP-based techniques on the user query improve
effectiveness
• Open issues
– No temporal evolution of content is considered
– Wikipedia is meant to be objective
01 Apr 2012
Gianluca Demartini, Claudiu S. Firan, Tereza Iofciu, Ralf Krestel, and Wolfgang Nejdl. Why Finding Entities in
Wikipedia is Difficult, Sometimes. In: "Information Retrieval" 13(5): 534-567, Springer, October 2010.
192

Time-Aware Entity Retrieval
• In some cases the time dimension is available
– News collections
– Blog postings
• News stories evolve over time
– Entities appear/disappear
– Analyse and exploit relevance evolution
– Decide about relevance at document level
• An Entity Search system can exploit the past to
find relevant entities
Gianluca Demartini, Malik Muhammad Saad Missen, Roi Blanco, Hugo Zaragoza. TAER: Time
Aware Entity Retrieval. CIKM 2010, Toronto, Canada.
01 Apr 2012
193

01 Apr 2012
194

• Evaluation
– P3, P5, AvgPrec
– Ties aware measures [McSherry and Najork, ECIR08]
• Paired t-test
– ** p<<0.01
– * p<0.05
• Related considered NonRelevant
01 Apr 2012
195

History Features
• We also tried
– Weight history features with doc length
– Weight history features with BM25
Feature P3 P5 MAP
F(e,d) .65 .56 .60
F(e,d1) .58 .53 .56
F(e,d-1) .64 .56 .62*
F(e,H) .66 .59** .66**
CoOcc(e,H) .62 .57 .65**
DF(e,H) .63 .57* .65**
01 Apr 2012
196

Dataset and Analysis
• TREC Novelty Track 2004
– 25 event topics
– 779 relevant news
• Entity annotations (7481 entities)
• Relevance judgements
• How useful is to find relevant sentences?
– P(e is Rel) 0.411 [0.404-0.417]
– P(e is NotRel) 0.168 [0.163-0.173]
– P(e is Rel | s is Rel) 0.547 [0.534-0.559]
– Sentences:
 21727 total 1.46 entity occurences
 5122 relevant 1.88 entity occurences
01 Apr 2012
197

Data Analysis
• How useful is looking at the past?
– P(e|d1) 0.893 [0.881-0.905]
– P(e|d-1) 0.701 [0.677-0.726]
• Is useful to consider sentence co-
occurence?
P(e1,e2) Relevant Related NotRelevant NotAnEntity
Relevant 0.24 0.08 0.03 0.07
Related 0.07 0.03 0.03
NotRelevant 0.07 0.05
NotAnEntity 0.04
01 Apr 2012
198

Approach
• Entity Ranking features for News articles
– Local Features
 F(e,d)
 FirstSenLen
 FirstSenPos
 Fsubj
 AvgBM25(q,s)
 SumBM25(q,s)
 History Features
• Feature combination
– Linear and Machine Learning
01 Apr 2012
199

Local Features
Feature P5 MAP
F(e,d) .56 .60
FirstSenLen .36 .45
FirstSenPos .31 .43
Fsubj .44 .50
AvgBM25(q,s) .30 .41
SumBM25(q,s) .44 .52
Feature P5 MAP
All Tied .34 .42
20001 Apr 2012

Is the past useful?
• Looking at previous documents
– Entity occurences so far F(e,H)
– Docs where the entity appeared so far
DF(e,H)
– Entity occurences in the previous doc
F(e,d-1)
– Frequency of entity the first time F(e,d1)
– Number of other entities with which the
entity co-occured so far CoOcc(e,H)
20101 Apr 2012

History Features
• * t-test p value < 0.05 as compared with F(e,d)
• ** t-test p value < 0.01 as compared with F(e,d)
Feature P5 MAP
F(e,d) .56 .60
F(e,d1) .53 .56
F(e,d-1) .56 .62*
F(e,H) .59** .66**
CoOcc(e,H) .57 .65**
DF(e,H) .57* .65**
20201 Apr 2012

0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 10 20 30 40 50 60
AvgPrec
i-th document (i.e., history size+1)
Using the History
Using the History
• Conclusion
– Evidence from past documents is very important
– Effectiveness should improve over time (run F(e,H))
01 Apr 2012 203

Discussion
• New search task: Time-Aware Entity
Retrieval
• Constructed evaluation benchmark
• Experimental Evaluation
– Investigated some features and
combinations
– Information from the past helps most
– Obtain 15% improvement over F(e,d)
20401 Apr 2012

Ranking Entities on the Web
• TREC Entity Track 2009-2010
– 50M web pages (including Wikipedia)
– Find related entities (return homepages)
01 Apr 2012
Entity Search on the Web 206

Ranking Entities on the Web
• Approaches
– Use Wikipedia (and infoboxes) as background info
– Extract entities from tables and lists
– Find the homepage given the entity name (see
ENS)
• Barack Obama -> www.barackobama.com
• Since 2010: 1 billion web pages
01 Apr 2012
207

Related Entity Finding
• Approaches
– Kaptein et al., CIKM10
• Exploits Wikipedia to improve entity retrieval
effectiveness
• Identifies entity types
• Wikipedia external links as source for entity homepage
• Anchor text index for entity search
– Bron et al, CIKM10
• Entity co-occurence
• Entity type filtering
• Context (relation type)
• Wikipedia-based homepage finding
01 Apr 2012
208

Discussion
01 Apr 2012
209

Expert Finding - Key Requirements
• Identify experts via self-nomination and/or
automated analysis of expert communications,
publications, and activities
• Classify the type and level of expertise of individuals
and communities
• Validate the breadth and depth of expertise of an
individual
• Recommend experts, including the ability to rank
order experts on multiple dimensions including skills,
experience, certification and reputation
01 Apr 2012
210

Current systems
• Hardly validate the breadth and depth of
expertise
– Count mentions
– Weight with relevance score
– Sometimes weight with authority of document
containing candidate mention
• Do not really attempt to classify the type and
level of expertise
01 Apr 2012
211

Evidence of Expertise
• Information about true expertise is often not
explicit in artifacts (as opposed to factual
knowledge)
• Information about expertise is expressed using
specialized terms and concepts
01 Apr 2012
212

How to improve?
• Integrate more sources of evidence
– CV information
– Project related data
• Including temporal information
– Training data (HR dept)
• Cost of achieving this evidence for expert vs.
non-expert as weighting factor
– Participation in TREC, authoring a book, getting a
PhD in IR, ...
Raymond D'Amore, Expert Finding in
Disparate Environments, 2008
01 Apr 2012
213

However...
• Two types of challenges to be overcome:
– System challenge
– Evaluation challenge
01 Apr 2012
214

System Challenges
• Multi-lingual entity extraction
• Privacy management
– E.g., Tacit can email top N experts with private
profiles (only recipient knows)
• Interoperability with heterogeneous data
sources
– IMAP, Exchange, Lotus Notes
– LDAP, JDBC/ODBC, XML repositories, Peoplesoft,
Oracle Financials, Word/Excel/PDF, ...
01 Apr 2012
215

Where is my data?
• > 80% of data not in relational databases
– Documents, spreadsheets, presentations
– Web pages
– Email, instant messages, news feeds
– Images, audio, video
01 Apr 2012
216

Dataspaces
• The complete set of information belonging to
one organization or task
• Examples:
– Personal dataspace
– Enterprise dataspace
– Community dataspace
• E.g., scientific, sports club, ...
“From Databases to Dataspaces: A New Abstraction for Information Management”,
Michael Franklin, Alon Halevy, David Maier, SIGMOD Record, December 2005.
01 Apr 2012
217

Conclusions so far...
• Expert finding could in principle use many
more resources that indicate expertise,
possibly more reliably, but it is difficult to
setup the research
– System challenges
– Data availability
• Motivates research in operational setting
– E.g., Raymond D'Amore, Expert Finding in
Disparate Environments, 2008
01 Apr 2012
218

Entity Search - Discussion
• Similar challenges as Expert Finding
• Entity information is spread over the Web
– In different formats (HTML, RDF, images)
– It is redundant (Wikipedia, DBPedia, homepage)
– It varies over time (e.g., population of a country)
– It is inconsistent (neutrino vs light speed)
01 Apr 2012
219

Entity Search - References
• Approaches exploit
– Wikipedia structure (links, categories)
• Kaptein et al., CIKM10 (REF)
• Demartini et al., IRJ 2010 (XER)
– Entity relations
• Bron et al., CIKM10 (REF)
• Demartini et al., CIKM10 (TAER)
– Graph-based methods
• Rode et al., INEX08 (XER)
• Iofciu et al., ECIR11 (XER)
– Probabilistic Models
• Weerkamp et al., INEX08 (XER)
• Balog et al., ECIR10 (XER)
01 Apr 2012
220

Entity Search - Discussion
• Structured data may be the way to improve
search effectiveness
– Entity identifiers
– Entity relations
01 Apr 2012
221

Introduction
• Unstructured or hybrid search over RDF data
– Supporting end-users
• Users who can not express their need in SPARQL
– Dealing with large-scale data
• Giving up query expressivity for scale
– Dealing with heterogeneity
• Users who are unaware of the schema of the data
• No single schema to the data
– Example: 2.6m classes and 33k properties in Billion Triples 2009
• Entity search
– Queries where the user is looking for a single entity named
or described in the query
– e.g. kaz vaporizer, hospice of cincinnati, mst3000
01 Apr 2012
223

Use cases in web search
Top-1 entity with
structured data
Related entities
Structured data
extracted from HTML
224

Architecture overview
Doc
1. Download, uncompress,
convert (if needed)
2. Sort quads by subject
3. Compute Minimal Perfect
Hash (MPH)
map
map
reduce
reduce
map reduce
Index
3. Each mapper reads part of the
collection
4. Each reducer builds an index
for a subset of the vocabulary
5. Optionally, we also build an
archive (forward-index)
5. The sub-indices are
merged into a single index
6. Serving
and
Ranking
1st part of the talk 2nd part
01 Apr 2012
225

Vertical index structure (reminder)
• One field (index) per property
• Positions are not required
• Query engine needs to support fields
 Dictionary is number of unique terms
 Occurrences is number of tokens
✗ Number of fields is a problem for merging, query performance
• In experiments we index the N most common properties
01 Apr 2012
226

BM25F Ranking
BM25(F) uses a term-frequency (tf) that accounts for the decreasing
marginal contribution of terms
where
vs is the weight of the field
tfsi is the frequency of term i in field s
Bs is the document length normalization factor:
ls is the length of field s
avls is the average length of s
bs is a tunable parameter
01 Apr 2012
227
Roi Blanco, Peter Mika, Sebastiano Vigna: Effective and Efficient Entity Search in RDF
Data. International Semantic Web Conference 2011:83-97

BM25F ranking cont.
• Final term score is a combination of tf and idf
where
k1 is a tunable parameter
wIDF is the inverse-document frequency:
• Finally, the score of a document D is the sum of
the scores of query terms q
01 Apr 2012 228

Hierarchical entity model
• Unstructured, structured and hierarchical entity model
• Hierrachical entity model
– Predicate type generation
– Predicate generation: importance of a predicate within its type
– Term generation: importance of a term is determined by the predicate
in which it occurs and all other predicates of that type in the entity
01 Apr 2012
229
Robert Neumayer, Krisztian Balog, and Kjetil Nørvåg. On the modeling of entities
for ad-hoc entity search in the web of data. In ECIR'12.

Query Independent Ranking
• The question is not which answer is more
relevant; i.e. all answers are relevant
• The task is finding out which of the answers
should be ranked higher
• Importance is subjective
• Closely related to the popularity of a resource?
01 Apr 2012
230
Lorand Dali, Blaz Fortuna, Thanh Tran Duc and Dunja Mladenic
Learning the Query-Independent Ranking of RDF Entity Search Results
In Proceedings of 9th Extended Semantic Web Conference (ESWC'12)

Towns from Andhra Pradesh
• Hyderabad
• Srisailam
• Chittoor
• Masulipatnam
• Chandavaram
• Mahbubnagar
• Gooty
• Vijaywada
• …
1. All answers are relevant
2. Ranking is important
3. Ranking is static
4. Hard to obtain the true ranking
01 Apr 2012
231

Learning to Rank
• Machine learning approach to building a ranking
model
• We know the true ranking (golden standard)
• We represent each answer (resource) as a feature
vector
• The final score is a linear combination of the features,
and the weights have to be learned
A
B
C
Pairwise preferences
A better than B
A better than C
B better than C
true ranking
01 Apr 2012
232

Ranking Features
• Importance derived
– from Graph analysis
– from Wikipedia
– from Web search engine
– from other external sources (N-gram databases)
01 Apr 2012
233

Graph Features
01 Apr 2012
234

Graph Features
• Pagerank
• Hubs and Authorities
• RDF graph features
– nRSubj - number of relations where this resource
appears as the subject
– nRObj - number of relations where this resource
appears as the object
– nLiteral - number of relations where this resource
appears as the subject and the object is a literal
01 Apr 2012
235

Importance of Wikipedia Pages
• Popularity
– How many people visited a particular page during
June-January 2010
– Data obtained from the Wikipedia access logs
available at: http://dammit.lt/wikistats
– Captures importance from the point of view of users
• Page length
– How much text a Wikipedia page contains
– Importance from the authors’ perspective
• Number of edits
– Importance from the editors’ perspective
01 Apr 2012
236

Features Approximating Importance
Correlate Well
• Compare rank based on page length and
based number of edits with page popularity
Spearman’s CC NDCG
Page length 0.60 0.84
Number of edits 0.78 0.93
01 Apr 2012
237

Web Search Features
• How many search results do we get in a web
search if we search for:
– The answer’s name
– Keywords from the answer’s description
• We used Yahoo! BOSS services to do the
search
• Querying the web for many resources is
expensive
01 Apr 2012
238

N-gram features
• Similar to web search features
• We look how many times the name of a
resource appears in a large N-gram database
(e.g. Google N-grams, Google Book N-grams,
etc.)
• A cheaper way to see how many times a
resource appears on the web or in books
01 Apr 2012
239

Introduction
• Intuitive keyword search interface over databases
• “A direction” of semantic search, which employs semantics of
– Relational information (structured data) in
– Different datasets
to produce complex structured, aggregated results to answer complex
information needs
• Short version of the Semantic Search tutorial at ESSIR’11
– Matching Techniques
– Ranking Techniques
• Complementary to DB keyword search tutorial, emphasizes
– The role of textual data: data graphs with textual content nodes
– Ranking
[Chen et al, SIGMOD09]
01 Apr 2012
241

Relational Entity Search
Matching

Structure
• Keyword search: keywords over data graphs
– Term matching
– Content matching
– Structure matching
• Schema-based keyword search
• Schema-agnostic keyword search
– Online search algorithms
– Index-based approaches
01 Apr 2012
243

Keyword search approaches
• Finding “substructures” matching keyword nodes
• Different result semantics for different types of data
– Textual data (Web pages connected via hyperlink)
– DB (tuple connected via foreign keys)
– XML (elements/attributes via parent-child edges)
• Commonly used results: Steiner tree / subgraph
– Connect keyword matching elements
– Contain one keyword matching element for every query keyword
– Minimal substructures: closely connected keyword nodes
• Query is ambiguous, lacks explicit structure constraints
– NP-hard, thus efficiency of matching is a problem
– Large amounts on candidate matches, thus ranking is a problem
01 Apr 2012
244

Keyword search on hybrid data graphs
Alice
Bob is a good friend
of mine. We went to
the same university,
and also shared an
apartment in Berlin in
2008. The trouble
with Bob is that he
takes much better
photos than I do:
trouble with bob
Bob
sunset.jpg
Beautiful
Sunset
Thanh
KIT
Germany
Semantic
Search
2009
Germany
PeterFluidOps 34
knows someone works at KITapartment shared Berlin Alice
Example information need
“Information about a friend of Alice, who shared an apartment with her in
Berlin and knows someone working at KIT.”
Term
matching
Content
matching
Structure
matching
01 Apr 2012
245

Term matching
• Distance-based (syntax)
– Levenshtein distance (edit distance)
– Hamming distance
– Jaro-Winkler distance
• Dictionary-based (semantics)
– Taxonomy
– Dictionary of similar words
– Translation memory
– Ontologies
01 Apr 2012
246

Content matching
• Retrieve partial matches
• Inverted list (inverted index)
ki  {< d1, pos, score, ...>,
< d2, pos, score, ...>, ...}
• Combine partial matches: union or join
shared
shared berlin alice= =
shared Berlin Alice shared Berlin Alice
D1 D1 D1
01 Apr 2012 247

Structure matching
• Retrieve structured data given patterns (e.g. triple patterns)
• Index on tables
• Multiple “redundant” indexes to cover different access patterns
• Combine: union or join
• Blocking, e.g. linear merge join (required sorted input)
• Non-blocking, e.g. symmetric hash-join
• Materialized join indexes
SP-index PO-index
=
=
=
?x ns:knows ?y. ?x ns:knows ?z.
?z ns: works ?v. ?v ns:name “KIT”
Per1 ns:works ?v ?v ns:name “KIT”
Per1 ns:works Ins1 Ins1 ns:name KIT
Per1 ns:works Ins1 Ins1 ns:name KIT
Structure not explicitly given in
query  exploration / other
kinds of join
01 Apr 2012
248

Structure
– Term matching
01 Apr 2012
249

Matching in keyword search – schema-based
Alice Bob KIT
• Operate on schema graph
• Query interpretation
– Compute queries instead of results
– Query presentation
– Query processing by DB engine
• Leverage the power of underlying DB query engine
Result 1
Result 2
[Tran et al, ICDE09]
[Hristidis et al, VLDB02] [Agrawal et al, ICDE02]
[Qin et al, SIGMOD09]
01 Apr 2012
250

Structure
– Term matching
01 Apr 2012
251

Matching in keyword search – schema-agnostic
Alice Bob KIT
• Operate on data graph
– No schema needed
– Flexibly support different types of data e.g. hybrid data
graphs
– Native tailored optimization
• Online in-memory graph search
• Using materialized indexes
Result 1
Result 2
[He et al, SIGMOD07] [Li et al, SIGMOD08]
[Tran et al, CIKM11]
[Kacholia et al, VLDB05]
01 Apr 2012 252

Online search – top-k exploration• Compute Steiner tree with distinct roots
• Backward expansion strategy
• Run Dijkstra’s single-source-shortest-path algorithms
– Explore shortest keyword-root paths
– To find root (an answer)
– Until k answers are found
– Approximate: no top-k guarantee, i.e. further answers found later from
other expansion paths may have higher score
• Complete top-k: terminate safely when lower bound of top-k
candidate is higher than upper bound of what can be achieved with
remaining inputs
[Bhalotia et al, ICDE02]
Alice Bob KIT
Result 1
01 Apr 2012 253

Taxonomy of matching approaches
• Schema-based vs. schema-agnostic
• Online search
– Complete top-k
– Approximate top-k
– Backward expansion, bidirectional search, undirected
subgraph exploration, dynamic programming
• Indexing for retrieval + join for combine
– Path retrieval, then path join
– Graph retrieval, then graph pruning
– Graph retrieval, then neighborhood / graph join
(neighborhood indexed as a set of paths)
01 Apr 2012
254

Relational Entity Search
Ranking

Structure
• Ranking paradigms
– Explicit model of relevance
– No notion of relevance
• Features
– Content-based
– Structure-based
– Structured-content-based
01 Apr 2012
256

Ranking paradigms
• No explicit notion of relevance: similarity between the
query and the document model
– Vector space model (cosine similarity)
– Language models (KL divergence)
• Explicit relevance model
– Foundation: probability ranking principle
– Ranking results by the posterior probability (odds) of being
observed in the relevant class:
)),...,(,),...,((),( ,,1,,1 qkqdtd wwwwCosdqSim 
)|(
)|(
log()|()||(),(
d
q
q
Vt
dq
tP
tP
tPKLdqSim


 

))|(1()|()|(  

DtDt
NtPRtPRDP
01 Apr 2012 257

Features
• Features are orthogonal to retrieval models
– Weights for query / document vectors?
– Language models for document / queries?
– Relevance models?
– What to use for learning to rank?
01 Apr 2012
258

Features
Dealing with ambiguities
• Content features
– Co-occurrences
• Terms K that often co-occur form a contextual
interpretation, i.e. topics (cluster hypothesis, distributional
semantics)
• “Berlin” and “apartment”  geographic context  Berlin as
city
– Frequencies: d more likely to be “about” a query term
k when d more often, mentions k (probabilistic IR)
• Structure features
– Structured-content-based: consider relevance at fine-
grained level of attributes
– Link-based popularity
– Proximity-based
Term
ambiguity
Content
ambiguity
Structure
ambiguity
01 Apr 2012
259

Content-based features – frequency
• Document statistics, e.g.
– Term frequency
– Document length
• Collection statistics, e.g.
– Inverse document frequency
– Background language models
)|()1(
||
)|( CtP
d
tf
tP d  
idf
d
tf
w dt 
||
,
• An object is more likely
about “Berlin”?
• When it contains a
relatively high number
of mentions of the
term “Berlin”
• When number of
mentions of term in
the overall collection is
relatively low
01 Apr 2012
260

Structure-based features – links
• PageRank
– Link analysis algorithm
– Measuring relative importance of nodes
– Link counts as a vote of support
– The PageRank of a node recursively depends on the
number and PageRank of all nodes that link to it
(incoming links)
• ObjectRank
– Types and semantics of links vary in structured data
– Authority transfer schema graph specifies connection
strengths
– Recursively compute authority transfer data graph
• An object (about “Berlin”) is more important?
• When a relatively large number of objects are linked to it
[Hristidis et al, TDS08]
How to incorporate it
into a content-based
retrieval model?
01 Apr 2012
261

• EASE, XRANK, BLINKS, etc.
• EASE
– Proximity between a pair of keywords
– Overall score of a JRT is aggregation on the score of keyword pairs
• XRANK
– Ranking of XML documents / elements
– Proximity of n is defined based on w, the smallest text window in n
that contains all search keywords
Structure-based features – proximity
• A structured result (e.g. Steiner tree) is more relevant?
• When it is more compact s.t. elements are closely related
[Li et al, SIGMOD08]
[Guo et al, SIGMOD03]
adopted from: [Chen et al, SIGMOD09]
How to incorporate it
into a content-based
retrieval model?
262

Structured-content-based model
• Consider structure of objects during content-based
modeling, i.e., to obtain structured content-based
model
– Content-based model for structured objects, structured
documents, database tuples…
)|()|( f
Ff
fd
d
tPtP  

• An object is more likely about “Berlin”?
• When its (important) fields / attributes contain a relatively
high number of mentions of the term “Berlin”
01 Apr 2012
263

P(w|Q) w
.077 palestinian
.055 israel
.034 jerusalem
.033 protest
.027 raid
.011 clash
.010 bank
.010 west
.010 troop
…
sample probabilities
  

UMM
k
i
ik MqPMwPMPqqwP
1
1 )|()|()()...,(
Palestinian
Israeli
raids
???
q1
q2
q3
w
M
M
M
)...(
)...,(
)...|()|(
1
1
1
k
k
k
qqP
qqwP
qqwPRwP 
Relevance model [Lavrenko et al, SIGIR01]
01 Apr 2012
264

Edge-specific relevance model
• Given a query Q={q1,…,qn}, a set of resources (FR) are retrieved
– E.g. Q={Hepburn, Holiday}, FR = {m1, p1, p4,m2, p2m2,m3}
• Based on FR results, an edge specific RMFR is constructed for each unique
edge e:
[Bicer et al, CIKM11]
01 Apr 2012
265

• Edge-specific resource model:
– Smoothing with model for the entire resource
• Use RM for query expansion: the score of a resource
calculated based on cross-entropy of edge-specific RMFR and
edge-specific RMr:
– Alpha allows to control the importance of edges
Edge-specific relevance model
01 Apr 2012
266

Ranking Steiner tree / join result tuples (JRT)
• Ranking aggregated JRTs:
– The cross entropy between the edge-specific RMFR (query model) and
geometric mean of combined edge-specific RMJRT:
• The proposed ranking function is monotonic with respect to the individual
resource scores (a necessary property for using top-k keyword search
algorithms)
01 Apr 2012
267

Taxonomy of ranking approaches
• Explicitly vs. non-explicitly relevance-based
• Different approaches for model construction
– Content-based ranking
– Structure-based ranking
– Structured-content-based ranking
01 Apr 2012
268

• Agrawal, S., Chaudhuri, S., and Das, G. (2002). DBXplorer: A system for keyword-based search
over relational databases. In ICDE, pages 5-16.
• Amer-Yahia, S. and Shanmugasundaram, J. (2005). XML full-text search: Challenges and
opportunities. In VLDB, page 1368.
• Bao, Z., Ling, T. W., Chen, B., and Lu, J. (2009). Effective xml keyword search with relevance
oriented ranking. In ICDE, pages 517-528.
• Bhalotia, G., Nakhe, C., Hulgeri, A., Chakrabarti, S., and Sudarshan, S. (2002). Keyword Searching
and Browsing in Databases using BANKS. In ICDE, pages 431-440.
• Bicer, V., Tran, T. (2011): Ranking Support for Keyword Search on Structured Data using
Relevance Models. In CIKM.
• Bizer, G., Lehmann, J., Kobilarov, G., Auer, S., Becker, C., Cyganiak, R., Hellmann, S. (2009):
DBpedia - A crystallization point for the Web of Data. J. Web Sem. (WS) 7(3):154-165
• Chen, Y., Wang, W., Liu, Z., Lin, X.(2009): Keyword search on structured and semi-structured
data. SIGMOD 2009:1005-1010
• Ding, B., Yu, J. X., Wang, S., Qin, L., Zhang, X., and Lin, X. (2007). Finding top-k min-cost
connected trees in databases. In ICDE, pages 836-845.
• Golenberg, K., Kimelfeld, B., and Sagiv, Y. (2008). Keyword proximity search in complex data
graphs. In SIGMOD, pages 927-940.
• Guo, L., Shao, F., Botev, C., and Shanmugasundaram, J. (2003). XRANK: Ranked keyword search
over XML documents. In SIGMOD.
• He, H., Wang, H., Yang, J., and Yu, P. S. (2007). BLINKS: Ranked keyword searches on graphs. In
SIGMOD, pages 305-316.
• Hristidis, V., Hwang, H., and Papakonstantinou, Y. (2008). Authority-based keyword search in
databases. ACM Trans. Database Syst., 33(1):1-40
01 Apr 2012
269

• Hristidis, V. and Papakonstantinou, Y. (2002). Discover: Keyword search in relational databases. In
VLDB.
• Kacholia, V., Pandit, S., Chakrabarti, S., Sudarshan, S., Desai, R., and Karambelkar, H. (2005).
Bidirectional expansion for keyword search on graph databases. In VLDB, pages 505-516.
• Kimelfeld, B. and Sagiv, Y. (2006). Finding and approximating top-k answers in keyword proximity
search. In PODS, pages 173-182.
• Ladwig, G., Tran, T. (2011): Index Structures and Top-k Join Algorithms for Native Keyword Search
Databases. In CIKM.
• Lavrenko, V. Croft, W.B. (2001): Relevance-Based Language Models. In SIGIR, pages 120-127.
• Li, G., Ooi, B. C., Feng, J., Wang, J., and Zhou, L. (2008). EASE: an effective 3-in-1 keyword search
method for unstructured, semi-structured and structured data. In SIGMOD.
• Liu, F., Yu, C., Meng, W., and Chowdhury, A. (2006). Effective keyword search in relational
databases. In SIGMOD, pages 563-574.
• Luo, Y., Lin, X., Wang, W., and Zhou, X. (2007). SPARK: Top-k keyword query in relational
databases. In SIGMOD, pages 115-126.
• Qin, L., Yu J. X., Chang, L. (2009) Keyword search in databases: the power of RDBMS. In SIGMOD,
pages 681-694.
• Tran, T., Herzig, D., Ladwig, G. (2011): SemSearchPro: Using Semantics throughout the Search
Process. In Journal of Web Semantics, 2011.
• Tran, T., Wang, H., Rudolph, S., Cimiano, P. (2009): Top-k Exploration of Query Graph Candidates
for Efficient Keyword Search on RDF. In ICDE.
• Vagelis Hristidis, L. G. and Papakonstantinou, Y. (2003). Efficient ir-style keyword search over
relational databases. In VLDB.
01 Apr 2012
270

Entity Search - Conclusions
• Historically:
• Expert Finding came first
• Generalized to Entity Search
– First on Wikipedia (easier)
– Then on the Web (harder)
• Over structured data
– AOR
– Relational Entity Search
01 Apr 2012
271

Entity Search - Conclusions
• What helps for Expert Finding?
– Good entity extraction
– Diverse evidence (Enterprise, web, user profiles)
• What helps for Entity Search?
– Wikipedia / DBPedia
– Entity types
– Anchors
01 Apr 2012
272

Next Steps for Entity Search
• Improve effectiveness for existing tasks
– Errors propagate along the pipeline
– Improve individual components (extraction,
linking, de-duplication, etc.)
– Improve ranking models by considering additional
evidence (as for expert finding)
• Work on top of Entity Search
– New User experiences (based on entities)
01 Apr 2012
273

Next Steps for Entity Search
• Novel entity-oriented search tasks
– Summary generation
– Entity Attribute Search (“At which age Nobel prize
winners in physics died?”)
– Entity Popularity (rank Nobel laureates by
popularity)
– Relation Search (“how are Michelle Obama,
Barack Obama, George Bush related?”)
– Event Search (“Jazz concerts in Barcelona”)
01 Apr 2012
274

Tutorial Outline
• Morning
– Data (Peter)
• Afternoon
01 Apr 2012 275

Outline
• TREC Enterprise
• INEX Entity Ranking
• TREC Entity
• INEX Data Centric
• SemSearch challenge
01 Apr 2012
277

Laboratory Evaluation in IR
• An IR test collection:
– A collection of documents
– A collection of search requests
– Relevance assessments for each search request
against the collection of documents
• Compare mean effectiveness (e.g., ratio of
relevant vs. non-relevant documents
retrieved) on ‘many’ search requests
01 Apr 2012
278

Relevance
TREC working definition:
If you were writing a report on the subject of
the topic and would use the information
contained in the document in the report, then
the document is relevant.
01 Apr 2012
279

Relevance – issues
• Binary relevance
– Regardless of how small the piece of relevant
information is in relation to the rest of the
document!
• Considered only in isolation
– Regardless of what other relevant information the
user has already seen
01 Apr 2012
280

Question:
• How well does this working definition of
relevance capture ‘real’ searcher behaviour?
– Ideal PhD student 
– Intelligence officer
– Medical doctor
– Patent lawyer
– My mum
01 Apr 2012
281

Question:
• What happens if we change the definition of
relevance to reflect different types of
searches?
01 Apr 2012
282

Entity Ranking
• Many users search for specific entities
instead of just any type of documents
01 Apr 2012
283

Ranking People
• Expert Finding in TREC-ENT (Enterprise Track)
• Collection:
– Corpus: crawl of *.w3.org sites
– People: names of 1092 people who may be experts
• Query:
– ‘information retrieval’
• Results:
– A list of people who know about information retrieval
01 Apr 2012
284

C-TEST
• TODO
• C-TEST
– http://es.csiro.au/C-TEST
01 Apr 2012
285

Expert Finding
01 Apr 2012
286

Evaluation Challenge
• How to get realistic data for research
evaluation?
• So far: build corpora from externally facing
parts of company intranets
01 Apr 2012
287

Evaluation: Enterprise Track
• Goal
– Build a test collection for Enterprise Search
• Corpora
– W3C corpus (2005, 2006)
– CSIRO corpus (2007)
• Tasks
– Email search (2005, 2006)
– Expert finding (2005-2007)
– Document search (2007)
01 Apr 2012
288

W3C Collection
• A crawl of the W3C
web, containing email
lists, wiki, web pages,
etc.
– 5.7 Gb crawl of w3.org
sites (in June 2004)
– 331,037 documents,
including ~175,000
emails
• A candidate list
01 Apr 2012
289

W3C Limitations
• Topics and assessments developed by
researchers instead of users
• Difficult to situate information needs in an
organization you are not part of –
representative of new members?
• Public e-mail lists; no private/personal
communication, and receivers unknown
01 Apr 2012
290

CERC
• Real enterprise
– CSIRO is the Commonwealth Scientific and
Industrial Research Organisation
– 50 sites in Australia and around the world
– 17 divisions, conducting research in areas such as
industrial physics, mining, and ICT
• Real users
– Science communicators
01 Apr 2012
291

CERC Limitations
• Real data, but...
• TREC has to use externally-facing data
– Publicly available pages from csiro.au only
– Can’t address security, databases, click data
• No candidate list!
– Participants do the assessments
01 Apr 2012
292

UvT Expert Collection
• Based on a publicly accessible database of UvT
employees who are involved in research or
teaching
• ~38,000 documents
– Research descriptions
– Course descriptions
– Publications
– Personal homepages
Balog et al., Broad Expertise Retrieval in Sparse
Data Environments, SIGIR 2007
01 Apr 2012 293

UvT Expert Collection
• Clean, heterogeneous, structured, and
focused (but comprises a limited number of
documents)
• Contains hierarchical information from an
organization (institutes and faculties)
• List of expertise areas provided by the
employees themselves
– http://ilk.uvt.nl/uvt-expert-collection/
Balog et al., Broad Expertise Retrieval in Sparse
Data Environments, SIGIR 2007
01 Apr 2012 294

‘In Situ’ Evaluation
Thomas and Hawking, Evaluation by comparing
result sets in context, CIKM 2006
01 Apr 2012 295

‘In Situ’ Evaluation
• Advantages
– Evaluation on live search system
– Click-through data useful in inferring preference,
even in the absence of explicit user judgements.
• Limitations
– Experimenter effect
– Experiments not repeatable
– A is better than B, but by how much? Multi-way
comparisons?
Thomas and Hawking, Evaluation by comparing result
sets in context, CIKM 2006
01 Apr 2012 296

Alternative benchmark model?
• Standard benchmark distributes the data +
queries + relevance information
• Q: Could we distribute the systems instead?
– Need organization(s) willing to evaluate in own
intra-net, with their own users
– Need participants with robust systems that run
off-site, unmonitored...
01 Apr 2012
297

INEX Entity Ranking
01 Apr 2012
298

Entity Ranking
• Topical query Q
• Entity (result) type TX
• A list of entity instances Xs
01 Apr 2012
299

Topic 60
Title
Entities
470 (dinghy) (#816578)
49er (dinghy) (#1006535)
Categories
dinghies (#30308)
Description
Narrative
470.
TX
Q
Xs
01 Apr 2012
300

INEX-XER
• INEX XML Entity Ranking Track
• Assumptions:
– Entities (Xs) are represented as Wikipedia pages
– Binary relevance
01 Apr 2012
301

Examples of Wikipedia Entities (TX)
• Art museums and galleries
• Countries
• Famous people
• Monarchs of the British Isles
• Artists
• Magicians
01 Apr 2012
302

Entity Ranking Scenarios (Q,TX)
• Impressionist art museums in Holland
• Countries with the Euro currency
• German car manufacturers
• Artists related to Pablo Picasso
• Countries involved in WWI
• Actors who played Hamlet
• Films shot in Venice
• Paul Auster novels
Many more examples on
http://www.ins.cwi.nl/projects/inex-xer/topics/
01 Apr 2012
303

Tasks
• Entity Ranking (ER)
– Given Q and TX, provide Xs
• List Completion (LC)
– Given Q and Xs[1..m]
– Return Xs[m+1..N]
01 Apr 2012
304

Topic 60
Title
Entities
470 (dinghy) (#816578)
49er (dinghy) (#1006535)
Categories
dinghies (#30308)
Description
Narrative
470.
01 Apr 2012
305

Easy Baseline?
//article[
about(Category:T and Q)]/title
01 Apr 2012
306

But not an easy task…
• Categories related to car companies:
– Peugeot_vehicles
– Citroën_vehicles
– Mercedes-Benz platforms
– DaimlerChrysler
– Motor vehicle manufacturers of Germany
– List_of_automobile_manufacturers
– Companies established in 1926
– …
01 Apr 2012
307

XML Structure to Rescue?
• Q = ‘James Bond’
• Xs = [Sean Connery, Roger Moore, …]
• Corpus data:
– Category:James Bond Actors
– Passage
Over 40 years after he first played the role, Connery is still widely regarded as the
definitive cinematic incarnation of James Bond, despite popular interpretations by the
likes of Roger Moore, Pierce Brosnan and Timothy Dalton, the latter often considered
closer to the Bond of the novels.
– Passage
To date, six actors have portrayed James Bond in the official series.
They are:
• Sean Connery (1962–1967; 1971)
• George Lazenby (1969)
• Roger Moore (1973–1985)
• Timothy Dalton (1987–1989)
• Pierce Brosnan (1995–2002)
• Daniel Craig (2006–present)
• So, should exploit XML Structure!!!
01 Apr 2012
308

Structure in the Collection
• Wikipedia contains semantically very rich
annotations:
– Wikipedia Categories
– Wikipedia Lists
– Times (1930, 1931, 1932, etc. etc.)
– Names
– Disambiguation pages
01 Apr 2012
309

So...
Recall our question:
• What happens if we change the definition of
relevance to reflect different types of
searches?
01 Apr 2012
311

INEX-XER Results
• XER Topics created by participants:
– Today, 38 genuine XER topics with completed
assessments
– ER: 8 groups, 18 runs submitted
– LC: 6 groups, 10 runs submitted
– Pool depth 50
01 Apr 2012
312

Ad-hoc vs. XER
• Many ad-hoc topics can serve as entity
ranking topics (but articles containing relevant
passages must be re-assessed)
– INRIA&RMIT created 30 training topics based on
the INEX 2006 ad-hoc collection
– XER Organizers created a similar set of 18 topics
based on the INEX 2007 ad-hoc collection
01 Apr 2012
313

Not relevant for XER…
• Articles on topic are not necessarily relevant
entities
– Actually, they are surprisingly often not!
• Articles that are or contain hubs (e.g., ‘list of
…’ pages) are not entities either, so they are
also not considered relevant
01 Apr 2012
314

Ad-hoc vs. XER
• On the 18 ad-hoc ’07 topics that were re-
assessed as XER topics, only about 35% out of
original relevant documents have been
assessed as relevant
– Most relevant Steven Seagul results are movies
(17/20); but most ‘French president in the 5th
republic’ results are not presidents or not
presidents in the 5th republic (6/129), and most
Bob Dylan songs are not songs (32/267)
01 Apr 2012
315

Ad-hoc vs. XER
• E.g., ‘Bob Dylan songs’ relevant articles/passages in ad-
hoc, but non-relevant entities in XER:
– Bob Dylan
– The Band
– Albums (however… what about track list?)
– Don’t look back (film)
– March on Washington for Jobs and Freedom (speech)
– 19XX [in music], 19XX grammy awards…
– Singer-songwriters that play covers/tributes/…
– Cities where Bob Dylan lived
Etc.etc.
01 Apr 2012
316

XER Close(r) to ‘Real’ User Need?
• Example: Ad-hoc topic 509, ‘songs of Bob
Dylan’
– Relevant passage…
The theme song is called This Wheel's on
Fire written by Bob Dylan and performed in
this instance by Julie Driscoll.
…however hidden in middle of lengthy page about
the ‘Absolutely fabulous’ TV series
• Clearly, the song’s article itself is the
preferable result, even for the ad-hoc task
01 Apr 2012
317

Perceived pool quality
• Participants missed only some entities they
know of…
– Toyota (‘hybrid cars’)
– Goat (‘herbivorous mammals’)
– Hillary Clinton (‘women in politics’)
• More of a problem for ad-hoc derived pools?
01 Apr 2012
318

Using hubs for validation
• Steven Seagal entity page
– Film-o-graphy:
• Prince of Central Park not in assessments
• Further web search: producer, not actor!
01 Apr 2012
319

Q: Non-topic-authors as judge?
Q: Use external information for judging?
01 Apr 2012 320

Q: Assessor’s Topic Knowledge not in
Collection?
01 Apr 2012
321

TREC (Web) Entity (Search)
01 Apr 2012
322

Related Entity Finding (REF)
• Topics:
– Input Entity:
Name + Homepage
– Target Type:
Person | Organisation | Product | Location
– Narrative:
Description of the relation in free text
01 Apr 2012
323

Entity Representation
(on the Web)
• Primary Entity Homepage:
devoted to and in control of the entity
– Gain = 3
• Relevant Entity Homepage:
devoted to the entity, but not in its control
– Gain = 1
• Not an Entity Homepage:
– Wikipedia Page
– News articles and blog posts, even if exclusively about the
entity
01 Apr 2012
324

The 2009 Pilot
• 20 topics
– Target Type:
Person | Organisation | Product
– 13 participants
01 Apr 2012
325

Example Topics
• Organizations that award Nobel prizes
• CDs released by the King's Singers
• Companies that John Hennessey serves on the
board of
• Universities that are members of the SEC
conference for football
• Scotch whisky distilleries on the island of Islay
01 Apr 2012
326

Lessons Learned
• Not that many entities in ClueWeb B
– Makes it difficult to define good topics, especially
product topics
• Wikipedia/DBPedia dominate approaches and
results
01 Apr 2012
327

2010
• ClueWeb A English
• 50 new REF topics
– Target Type:
Person | Organisation | Product | Location
– 47 topics judged, pool depth 20
• 20 2009 topics re-assessed
• Wikipedia pages not accepted as valid answers,
primary homepages rewarded more
– Evaluation script:
http://trec.nist.gov/data/entity/10/eval-entity.pl
• 14 participants
01 Apr 2012
328

2011
• Modifications to simplify the evaluation and to
improve cross-system comparison
– Binary relevance (only Primary HP)
– Groups using Web Search Engines should also submit a run
using the Lemur ClueWeb Online Query Service
• 50 new REF topics
– Target Type not limited (extracted from Narrative)
– Support document per answer (could be WP page)
• 10 participants (only 4 doing REF )
01 Apr 2012
329

Entity List Completion (2010)
• BTC-2009 – Billion Triple Collection
– LOD sample, 1.14B statements, 17GB zipped
• 13 topics
– Out of 20 REF 2009, skipping topics without
results or without example entities
– Mapped type to most specific class in DBPedia
ontology
– Mapped known relevant entity HPs to URIs in BTC-
2009
01 Apr 2012
330

ELC Assessments (2010)
• By track coordinators
(yes, I have read my share of the semantic web )
• Many answers of correct type, but not in
correct relation to example entities (and,
many descriptors of input entity had been
retrieved)
• 5 topics without relevant entities, 1 topic “un-
judgeable”
01 Apr 2012
331

“Semantic” Sub-tasks 2011
• Sindice-2011 dataset
– Various Linked Data RDF dumps
– 1.7B entities, 11B RDF statements (230M docs)
– Web documents containing RDF, RDFa and
Microformats documents (crawled by the Sindice
Bot, since 2009)
http://data.sindice.com/trec2011/
S. Campinas, D. Ceccarelli, T. E. Perry, R. Delbru, K. Balog, and G.
Tummarello. The Sindice-2011 Dataset for Entity-Oriented Search in the
Web of Data. In EOS, SIGIR 2011 workshop, July 28, Beijing, China.
01 Apr 2012
332

“Semantic” Sub-tasks 2011
• REF-LOD
– Like REF, but results are URIs in Sindice instead of
docids in ClueWeb
• But, not a re-usable collection – only 1 participant
• Entity List Completion
– Like ELC 2010, but examples not given as URI, but
only as ClueWeb identifier
01 Apr 2012
333

Meanwhile at SemEval...
01 Apr 2012
334

Web People Search
• WePS1 (2007) focused on Disambiguation of person
names in search results
– determine #people for that person name
– assign documents to each referent
• WePS2 (2009) introduced Attribute Extraction
– 18 attribute types including birthdate and –place,
occupation, affiliation, ...
• WePS3 (2010) introduced disambiguation for
organization names (motivated by Online Reputation
Management)
http://nlp.uned.es/weps/01 Apr 2012

Other Entity Oriented Evaluations
• Where to go, now that TREC Entity is “on a
break”?!
01 Apr 2012
336

TREC Knowledge Base Acceleration
• Given
– Incoming text stream (news and social media
content)
• First month w/ human-generated labels as training data
– A target entity from a knowledge base (for now:
people, specified by their Freebase and Wikipedia
entries)
• Score each item (“document”) based on how
“pertinent” it is to the target KB node
01 Apr 2012
337

TAC Knowledge Base Population
• Tasks related to extracting information about entities with
reference to an external knowledge source (Wikipedia
infoboxes)
• KBP 2011 had three tasks:
– entity-linking: given an entity name (person, organization, or
geopolitical entity) and a document containing that name, determine
the KB node for that entity or add a new node for the entity if it is not
already in the KB
– slot-filling: given a named entity and a pre-defined set of attributes
(“slots”) for the entity type, augment a KB node for that entity by
extracting all new learnable slot values from a large corpus of
documents
– temporal slot-filling: similar to the regular slot-filling task, but also
requests time intervals to be specified for each extracted slot value.
01 Apr 2012
338

CLEF RepLab
• Online reputation management (ORM)
– Continuation of the WePS3 ORM task
– Two tasks, evaluated on Twitter data:
• Monitoring task: thematically cluster tweets including a
company’s name
• Profiling task: annotate tweets according to their
polarity
01 Apr 2012
339

INEX Data-Centric Track 2011
• IMDB data collection
– XML generated from plain text files published on
the IMDB web site on April 10, 2010
– Movies and persons involved in movies, e.g.
actors/actresses, directors, producers etc.
• ~4M objects, including 1.5M movies and 1.8M actors
• Each object is richly structured
– Each movie has title, rating, directors, actors, plot, keywords,
genres, release dates, trivia, etc.
– Each person has name, birth date, bio, filmography, etc.
01 Apr 2012
340

• Ad-hoc
– Return a ranked list of results, i.e. objects, or equivalently
documents in the IMDB collection, estimated relevant to
the user’s information need
– Known-item, list and informational topics
• Faceted search task
– Recommend appropriate facet-values for the user to refine
the query and thus quickly identify what he/she really
wants in the large set of results.
– A User Simulation System (USS) interacts with a so-called
FSM to simulate a faceted search session; using simulated
interaction cost as evaluation measure (alongside NDCG of
facet values)
01 Apr 2012
341

• 2012 uses DBPedia and Wikipedia
01 Apr 2012
342

Evaluation
Ad-hoc Object Retrieval

Semantic Search Evaluation
• Started at SemSearch 2010
• Two tasks
– Entity Search
• Queries where the user is looking for a single real world
object
• Pound et al., Ad-hoc Object Retrieval in the Web of
Data, WWW 2010.
– List search (new in 2011)
• Queries where the user is looking for a class of objects
01 Apr 2012
344

Semantic Search Evaluation
• Billion Triples Challenge 2009 dataset
• Evaluated using Amazon’s Mechanical Turk
– See Halpin et al. Evaluating Ad-Hoc Object
Retrieval, IWEST 2010
• Prizes sponsored by Yahoo! Labs (2011)
01 Apr 2012
345

SemSearch Entity Search Track
• Retrieval of data related to a single entity
• 50 queries explicitly naming an entity(possibly w/
context)
– Selected from Search Query Tiny Sample v1.0 (i.e., Yahoo!
Webscope): web search queries sampled from US query
log of January, 2009.
– Queries asked by at least three different users; queries
with long number sequences removed (privacy reasons)
• 2011: same setup, but queries derived from a mix of
Microsoft and Yahoo! Logs
01 Apr 2012
346

SemSearch List Query Track
• List queries
–Queries that describe a set of entities
–The answer is a closed set
–Relatively small number of possible
answers
–The answer is not likely to change
01 Apr 2012
347

SemSearch List Query Track
• Hand-picked but not hand-written
–Yahoo! Search logs
• Queries from the Tiny Sample v1.0
dataset
• Queries with clicks on Wikipedia
–TrueKnowledge
• Recent queries
01 Apr 2012
348

Data set
• Billion Triples Challenge 2009 data set
–Blank nodes encoded as URIs
• A data set combining crawls of multiple
semantic search engines
–doesn’t necessarily match the current state
of the Web
–doesn’t necessarily match the coverage of
any particular search engine
http://km.aifb.uni-karlsruhe.de/ws/dataset_semsearch2010/000-CONTENTS01 Apr 2012

Collecting the results
• Submissions via semsearch.yahoo.com
– NTNU, IIIT Hyderabad, DERI, U of Delaware, Daiict
– max. 3 submissions per team per track
• Pooling of results
– Top 20 results are evaluated
– Despite validation, still problems
• e.g. N-Triples encoded URIs, lowercased URIs
• Collecting triples for each result
– All triples where the URI is the subject
– Discarded URIs that didn’t appear as subject
• Rendering result display
– Values are clipped at 300 chars (last # or / for object-properties)
– RDF built-ins shown first
– Preference to English language values
01 Apr 2012
350

Assessment with Amazon Mechanical
Turk
– Evaluation using non-expert judges
• Paid $0.2 per 12 results
– Typically done in 1-2 minutes
– $6-$12 an hour
– Sponsored by the European SEALS project
• Each result is evaluated by 5 workers
– Workers are free to choose how many
tasks they do
• Makes agreement difficult to compute
0
50
100
150
200
250
300
Number of tasks completed per worker (2010)
352

Evaluation form
01 Apr 2012
353

Evaluation form
01 Apr 2012
354

Catching the bad guys
– Payment can be rejected for workers who try to
game the system
• An explanation is commonly expected, though cheaters
rarely complain
– We opted to mix control questions into the real
results
• Gold-win cases that are known to be perfect
• Gold-loose cases that are known to be bad
– Metrics
• Avg. and std. dev on gold-win and gold-loose results
• Time to complete

Catching the bad guys
01 Apr 2012
356
Worker Known
bad
Real Known Good Total
N
Time to
complete
(sec)N Mean N Mean N Mean
badguy 20 2.556 200 2.738 20 2.684 240 29.6
goodguy 13 1 130 2.038 13 3 156 95
whoknows 1 1 21 1.571 2 3 24 83.5

Lessons learned
• Finding complex queries is not easy
– Query logs from web search engines contain simple queries
– Semantic Web search engines are not used by the general public
• Computing agreement is difficult
– Each judge evaluates a different number of items
• Result rendering is critically important
• The Semantic Web is not necessarily all DBpedia
– sub30-RES.3 40.6% DBpedia
– sub31-run3 93.8% Dbpedia
• Follow up experiments validated the Mechanical Turk approach
– Blanco et al. Repeatable and Reliable Search System Evaluation using
Crowd-Sourcing, SIGIR2011
01 Apr 2012
357

SemSearch Next Steps
• Achieve repeatability
–Simplify our process and publish our tools
–Automate as much as possible… except the
Turks ;)
• Web site for evaluation
–Continuous submission?
• Join the discussion at
semsearcheval@yahoogroups.com
01 Apr 2012
358

Entity Search Evaluation
• What’s next?
– Continuous evaluation for AOR
– INEX Data-Centric Track 2012
– Knowledge Base Acceleration Track at TREC 2012
– New entity-centric search tasks?
– New data collections?
01 Apr 2012
359

01 Apr 2012
360

From Expert Finding to Entity Search on the Web

More Related Content

What's hot (20)

Similar to From Expert Finding to Entity Search on the Web (20)

Recently uploaded (20)

From Expert Finding to Entity Search on the Web

Editor's Notes